Meeting Title: ABC Project Architecture Discussion Date: 2025-12-12 Meeting participants: Mustafa’s Loom Notetaker, Mustafa Raja, Samuel Roberts
WEBVTT
1 00:00:16.110 ⇒ 00:00:17.200 Samuel Roberts: B.
2 00:00:20.290 ⇒ 00:00:21.250 Mustafa Raja: Hey, how are you?
3 00:00:21.730 ⇒ 00:00:25.359 Samuel Roberts: Doing alright. A lot of meetings… haven’t been…
4 00:00:25.360 ⇒ 00:00:26.040 Mustafa Raja: Yeah.
5 00:00:26.850 ⇒ 00:00:32.999 Samuel Roberts: Haven’t been able to look at the code for the ABC stuff yet, but, so, I thought you…
6 00:00:33.000 ⇒ 00:00:33.760 Mustafa Raja: Silo.
7 00:00:34.250 ⇒ 00:00:38.899 Samuel Roberts: Villa’s good! We had a meeting with them today. The code is a little… it’s weird, they gave us…
8 00:00:39.200 ⇒ 00:00:45.090 Samuel Roberts: Some repos, but, like, There was a hidden branch on one of them that had updated code.
9 00:00:45.800 ⇒ 00:00:49.369 Samuel Roberts: And so, we’re not sure if we have the right code for the other repos.
10 00:00:49.590 ⇒ 00:00:51.119 Samuel Roberts: It’s just… it’s weird.
11 00:00:51.120 ⇒ 00:00:55.899 Mustafa Raja: I’m hearing about hidden branches for the first time.
12 00:00:55.900 ⇒ 00:00:59.410 Samuel Roberts: I don’t know, yeah, like, it didn’t show up there until I, like…
13 00:00:59.690 ⇒ 00:01:03.350 Samuel Roberts: fetched, and then it found the remote branch, and I don’t know what was going on there.
14 00:01:03.690 ⇒ 00:01:05.990 Samuel Roberts: And then I re-downloaded it to see if it was…
15 00:01:06.520 ⇒ 00:01:09.369 Samuel Roberts: What it was, and it did the same thing, so it was really odd.
16 00:01:10.890 ⇒ 00:01:11.600 Mustafa Raja: Yeah.
17 00:01:12.880 ⇒ 00:01:14.509 Samuel Roberts: But, yeah, it’s alright.
18 00:01:14.750 ⇒ 00:01:16.300 Samuel Roberts: Abc.
19 00:01:16.300 ⇒ 00:01:20.070 Mustafa Raja: Yeah, yeah, let’s, let’s… Oh, sorry, what’s…
20 00:01:20.850 ⇒ 00:01:21.799 Samuel Roberts: What were you saying?
21 00:01:22.200 ⇒ 00:01:25.870 Mustafa Raja: Yeah, I just had one question. What database are we using with Lilo?
22 00:01:26.650 ⇒ 00:01:28.380 Samuel Roberts: It’s gonna be Postgres.
23 00:01:28.990 ⇒ 00:01:29.710 Mustafa Raja: On railways.
24 00:01:29.710 ⇒ 00:01:33.849 Samuel Roberts: Probably just on Railway, yeah, because that’s how we have this repo set up, but…
25 00:01:35.140 ⇒ 00:01:39.479 Samuel Roberts: That way it’s all in one place for them, rather than giving them, like, superbase and railway.
26 00:01:40.030 ⇒ 00:01:41.240 Mustafa Raja: Yeah, makes sense.
27 00:01:42.240 ⇒ 00:01:45.400 Samuel Roberts: It is nice to have the Superbase, like, UI and stuff, but…
28 00:01:46.080 ⇒ 00:01:47.000 Samuel Roberts: I don’t know if we’re gonna.
29 00:01:47.000 ⇒ 00:01:54.779 Mustafa Raja: My, my, my selling point, or my, favorite thing about Superbase is auth.
30 00:01:55.650 ⇒ 00:01:56.799 Samuel Roberts: Yeah, and we’re…
31 00:01:56.800 ⇒ 00:01:57.240 Mustafa Raja: gets…
32 00:01:57.240 ⇒ 00:01:58.600 Samuel Roberts: with… yeah.
33 00:01:59.870 ⇒ 00:02:11.260 Mustafa Raja: Yeah, so we don’t have to, you know, worry about sessions, we don’t have to worry about, tables and stuff, and yeah, it’s only that. Other than that, yeah, pretty good.
34 00:02:12.150 ⇒ 00:02:15.569 Samuel Roberts: Yeah, we’ve got a better offset up with that.
35 00:02:15.820 ⇒ 00:02:19.920 Samuel Roberts: Which is… seems like a pretty nice package, like, open source package.
36 00:02:19.920 ⇒ 00:02:21.099 Mustafa Raja: Yeah, I’ve worked with it.
37 00:02:21.470 ⇒ 00:02:25.649 Samuel Roberts: Okay, yeah, I haven’t… I had touched Next off on another project, but not better off yet, so…
38 00:02:25.650 ⇒ 00:02:37.829 Mustafa Raja: Yeah, I had to… when I used to freelance, I had to, you know, move a lot of stuffs, a lot of the stuff that I would work on would… would… someone else would have worked on already, so…
39 00:02:37.830 ⇒ 00:02:38.310 Samuel Roberts: Yup.
40 00:02:38.310 ⇒ 00:02:48.319 Mustafa Raja: That is how I learned most of my tools. So this is, I moved it back to, what’s it called? A photo, and this is…
41 00:02:48.320 ⇒ 00:02:48.830 Samuel Roberts: Okay.
42 00:02:48.830 ⇒ 00:02:53.430 Mustafa Raja: This is exactly the same way Andy would respond.
43 00:02:53.590 ⇒ 00:02:59.000 Mustafa Raja: So yeah, it was just the model thing. Okay, good.
44 00:02:59.000 ⇒ 00:02:59.540 Samuel Roberts: Good.
45 00:03:00.130 ⇒ 00:03:04.550 Mustafa Raja: 5… 5.1 would… is a lot different, yes.
46 00:03:04.550 ⇒ 00:03:05.600 Samuel Roberts: Yeah…
47 00:03:05.600 ⇒ 00:03:15.229 Mustafa Raja: Part of the, part of the reason there is maybe, the, you know, context tokens, you know? .
48 00:03:15.230 ⇒ 00:03:15.820 Samuel Roberts: Right.
49 00:03:15.820 ⇒ 00:03:26.660 Mustafa Raja: is a lot less, and 5, or we can say as the generation is growing, the context window is, becoming larger and larger, so…
50 00:03:26.840 ⇒ 00:03:35.679 Mustafa Raja: Part of the reason, I think, 4O responds this way is to, you know, conserve tokens or something, you know?
51 00:03:35.850 ⇒ 00:03:41.799 Samuel Roberts: Yeah, I think so. Yeah, I mean, it could be worth looking at changing at some point, but I didn’t want to mess with that yet.
52 00:03:41.800 ⇒ 00:03:43.320 Mustafa Raja: Yeah, but, yeah, yeah.
53 00:03:44.180 ⇒ 00:03:45.500 Samuel Roberts: Get this going, but… okay.
54 00:03:45.500 ⇒ 00:03:57.589 Mustafa Raja: Yeah, this looks good to me. But, one thing I would say is this is on our end right now, so what do I mean by that? We are using our,
55 00:03:57.720 ⇒ 00:04:09.889 Mustafa Raja: Azure, models, right? And I know that ABC already has deployed the models on their side too, but I could not find
56 00:04:10.210 ⇒ 00:04:12.860 Mustafa Raja: The keys, you know, to the…
57 00:04:12.860 ⇒ 00:04:13.190 Samuel Roberts: Okay.
58 00:04:13.190 ⇒ 00:04:25.889 Mustafa Raja: what’s it called? Endpoints. So, yeah, for now, it’s all us. But, I guess once Casey is back, I believe he should have something.
59 00:04:26.120 ⇒ 00:04:31.280 Mustafa Raja: Since this was set up, in Editon before me, you know? So…
60 00:04:31.280 ⇒ 00:04:35.860 Samuel Roberts: Yeah, yeah, so hopefully he has those. If not, we can make sure to get them before we deploy for theirs.
61 00:04:36.390 ⇒ 00:04:39.160 Samuel Roberts: Yup. But I think I’m not too worried about that yet.
62 00:04:39.980 ⇒ 00:04:44.979 Mustafa Raja: Yeah, so what… what should the next steps be, really, now?
63 00:04:44.980 ⇒ 00:04:46.260 Samuel Roberts: Okay, well…
64 00:04:46.260 ⇒ 00:04:47.530 Mustafa Raja: How do you want to test it?
65 00:04:48.140 ⇒ 00:04:51.340 Samuel Roberts: I tested it a little… I haven’t tested the new model, you’re right, I forgot about that.
66 00:04:53.890 ⇒ 00:04:56.190 Samuel Roberts: I… Well, here… okay.
67 00:04:56.480 ⇒ 00:04:59.930 Samuel Roberts: Let me talk you through what my thought process was the other day before I…
68 00:05:00.270 ⇒ 00:05:04.100 Samuel Roberts: I had to sign off, because I wasn’t thinking straight.
69 00:05:05.200 ⇒ 00:05:12.220 Samuel Roberts: I’m wondering now, with Maestra, if we want to go about it in a different… architecture.
70 00:05:12.650 ⇒ 00:05:17.910 Samuel Roberts: So… Like, right now, this is an agent that has tools, right?
71 00:05:19.780 ⇒ 00:05:20.410 Mustafa Raja: Yes.
72 00:05:20.410 ⇒ 00:05:27.270 Samuel Roberts: But we could potentially do, like, a number of agents built into a workflow, that…
73 00:05:27.980 ⇒ 00:05:30.280 Samuel Roberts: Could all have tools available to them.
74 00:05:30.930 ⇒ 00:05:31.779 Samuel Roberts: And the reason I…
75 00:05:31.780 ⇒ 00:05:35.370 Mustafa Raja: Something like, something that we did for Interlude.
76 00:05:36.550 ⇒ 00:05:41.569 Samuel Roberts: Kind of, yeah. Where… and it’s even similar to the routing agent that we had, but potentially making it, like, a…
77 00:05:41.840 ⇒ 00:05:50.039 Samuel Roberts: Let me… let me actually send you what I had done. Let me get it open here. I don’t think I have it open anymore, but I was chatting with ChatGPT,
78 00:05:50.190 ⇒ 00:05:55.450 Samuel Roberts: about… Other potential architectures.
79 00:05:56.090 ⇒ 00:05:58.819 Samuel Roberts: And I think I was looking at…
80 00:06:00.560 ⇒ 00:06:02.549 Samuel Roberts: Yeah, let me, let me just share this…
81 00:06:02.550 ⇒ 00:06:03.240 Mustafa Raja: Sweet.
82 00:06:03.240 ⇒ 00:06:04.920 Samuel Roberts: No, no, no, I can right now, though, hold on.
83 00:06:06.810 ⇒ 00:06:07.390 Mustafa Raja: Okay.
84 00:06:10.170 ⇒ 00:06:14.990 Samuel Roberts: B… I can even send this to you, too, just to… Ugh.
85 00:06:15.160 ⇒ 00:06:18.229 Samuel Roberts: TriTPT has been really slow for me, I don’t know if you’ve…
86 00:06:19.340 ⇒ 00:06:23.289 Samuel Roberts: I feel like the actual interface for ChatGPT is…
87 00:06:24.910 ⇒ 00:06:28.150 Mustafa Raja: Do you use the web one, or do you have the app installed?
88 00:06:28.690 ⇒ 00:06:31.420 Samuel Roberts: I have the web one, but I think it’s probably the same thing, isn’t it?
89 00:06:32.910 ⇒ 00:06:38.009 Mustafa Raja: For me, it really was not. The app one works better for me.
90 00:06:38.250 ⇒ 00:06:41.329 Samuel Roberts: Oh, really? Okay, maybe I’ll have to try it, I just assumed it was just a rapper, but…
91 00:06:44.180 ⇒ 00:06:54.580 Samuel Roberts: Let me, let me share this real quick, while I’m talking. Where is… We’ll share… So, I had…
92 00:06:55.100 ⇒ 00:07:01.129 Samuel Roberts: let me just go up to the top of this one, but basically I took the output of N8N, like the JSON, and was just like…
93 00:07:02.010 ⇒ 00:07:04.880 Samuel Roberts: I wanna move this, let me get this thing out of here.
94 00:07:06.580 ⇒ 00:07:09.429 Samuel Roberts: focus on the flow that… and then I actually went in and…
95 00:07:09.930 ⇒ 00:07:13.220 Samuel Roberts: It was looking at too much stuff, because there’s a lot in that N8N.
96 00:07:13.630 ⇒ 00:07:18.319 Samuel Roberts: So I went in, copied it, deleted everything that wasn’t part of the main ANDI.
97 00:07:18.780 ⇒ 00:07:24.200 Samuel Roberts: And then… I just wanted to understand what we were trying to do. I wasn’t really looking for it to do all this, where it was like…
98 00:07:24.900 ⇒ 00:07:30.559 Samuel Roberts: doing it out, a more faithful architecture. But really what was good here was when I said…
99 00:07:33.210 ⇒ 00:07:36.040 Samuel Roberts: It generated so much, so let’s dive into option 2.
100 00:07:36.820 ⇒ 00:07:39.189 Samuel Roberts: Yes, okay, so these are the options here.
101 00:07:40.230 ⇒ 00:07:46.460 Samuel Roberts: I said, can you brainstorm some other ideas for a high-level structure? So one was a super agent with different workflows.
102 00:07:46.840 ⇒ 00:07:48.720 Samuel Roberts: Which is kind of what we have right now.
103 00:07:50.310 ⇒ 00:07:57.239 Samuel Roberts: The other one was a multi… multi-expert agents and router, which is kind of what we… we had a router, but it was just still, like…
104 00:07:57.490 ⇒ 00:07:59.319 Samuel Roberts: To the same agent, right?
105 00:07:59.670 ⇒ 00:08:00.510 Mustafa Raja: Yeah.
106 00:08:00.830 ⇒ 00:08:10.430 Samuel Roberts: So I think the idea, and this is what I was kind of thinking would be good for Maestra, is that this would let us point different agents to the right
107 00:08:10.730 ⇒ 00:08:12.130 Samuel Roberts: Data sources?
108 00:08:15.580 ⇒ 00:08:16.320 Samuel Roberts: Lum.
109 00:08:16.320 ⇒ 00:08:20.690 Mustafa Raja: This would mean… so we have two source of truths.
110 00:08:21.200 ⇒ 00:08:22.060 Mustafa Raja: Right.
111 00:08:22.350 ⇒ 00:08:33.509 Mustafa Raja: And maybe 3, to be honest, because, I see some data directly ingested into, Bronx, so let’s consider that data also a source of truth, right? So we have 3.
112 00:08:33.510 ⇒ 00:08:34.159 Samuel Roberts: Alright.
113 00:08:34.169 ⇒ 00:08:39.409 Mustafa Raja: One is going to be the docs.
114 00:08:39.719 ⇒ 00:08:44.719 Mustafa Raja: Centradocs, and then, and then, what’s it called?
115 00:08:45.079 ⇒ 00:08:46.469 Mustafa Raja: the zip codes?
116 00:08:46.749 ⇒ 00:08:59.249 Mustafa Raja: The database stuff that, the text of SQL that Casey built, and then whatever the prompt has, which… which I think is just cancellation flows, and that takes up 75% of
117 00:08:59.409 ⇒ 00:09:02.199 Mustafa Raja: The prompt, which is not good, right? So, anyway…
118 00:09:02.200 ⇒ 00:09:02.570 Samuel Roberts: So that’s when.
119 00:09:02.570 ⇒ 00:09:04.610 Mustafa Raja: back to this? Yeah, that’s where.
120 00:09:04.610 ⇒ 00:09:09.500 Samuel Roberts: I’m thinking, like, the cancellation flows, we could make its own… Knowledge base.
121 00:09:09.750 ⇒ 00:09:10.959 Mustafa Raja: Yeah, so do you think…
122 00:09:12.010 ⇒ 00:09:23.379 Mustafa Raja: Do you think… so this best knowledge agent, what I think, here it’s trying to say is the central docs, right?
123 00:09:23.380 ⇒ 00:09:30.380 Samuel Roberts: It didn’t… it didn’t quite know what to call it, but… and these might not be the exact agents, and this is just… I was trying to get it to give me an idea here.
124 00:09:30.560 ⇒ 00:09:35.660 Samuel Roberts: But so, the other ones it had was Retrieved and Generate, which was interesting.
125 00:09:36.630 ⇒ 00:09:38.080 Mustafa Raja: Retrieve, then generate.
126 00:09:39.470 ⇒ 00:09:42.839 Samuel Roberts: Which I don’t fully know how it would do the retrieval.
127 00:09:43.320 ⇒ 00:09:52.459 Mustafa Raja: Run template queries on this… Combine, chunky, chunks, contact… but this… isn’t this the way it just works, though?
128 00:09:53.070 ⇒ 00:10:03.750 Samuel Roberts: Well, it does the lookup and then the… I don’t know, this is a weird one, I didn’t fully understand that, and then Tool Orchestrated seemed more complicated, like, different, multi-agent team with supervisor is another one, but that seems…
129 00:10:03.750 ⇒ 00:10:09.329 Mustafa Raja: Yeah, this is the… this is something that we might have done for… for, Interlude Rick.
130 00:10:09.330 ⇒ 00:10:10.100 Samuel Roberts: Yeah, I think.
131 00:10:10.100 ⇒ 00:10:12.679 Mustafa Raja: We had multiple agents and then a supervisor, yeah.
132 00:10:12.680 ⇒ 00:10:17.429 Samuel Roberts: Yeah, so I think, I think this, like, 5 and 2…
133 00:10:17.650 ⇒ 00:10:19.330 Samuel Roberts: were kind of what I was thinking.
134 00:10:20.090 ⇒ 00:10:20.820 Mustafa Raja: Yeah.
135 00:10:21.200 ⇒ 00:10:21.980 Mustafa Raja: Yeah.
136 00:10:21.980 ⇒ 00:10:26.560 Samuel Roberts: They’re similar, multi-expert agents and, you know, the other one… what was it?
137 00:10:26.900 ⇒ 00:10:28.160 Samuel Roberts: supervisor.
138 00:10:28.600 ⇒ 00:10:30.120 Mustafa Raja: Yeah, yeah, so…
139 00:10:30.120 ⇒ 00:10:32.689 Samuel Roberts: Seems a little overkill for what we need right now, I think.
140 00:10:34.420 ⇒ 00:10:35.120 Mustafa Raja: Yes.
141 00:10:35.120 ⇒ 00:10:35.500 Samuel Roberts: I think…
142 00:10:35.500 ⇒ 00:10:35.990 Mustafa Raja: not…
143 00:10:35.990 ⇒ 00:10:41.080 Samuel Roberts: The idea here is, yeah, overkill of traffic is low, there you go, more expensive to run.
144 00:10:41.950 ⇒ 00:10:47.660 Samuel Roberts: So I think, I think I was, I was leaning towards… oh, here we go.
145 00:10:47.790 ⇒ 00:10:49.769 Samuel Roberts: I was leaning towards this…
146 00:10:50.960 ⇒ 00:10:55.550 Samuel Roberts: And so the idea… and, like, the only con here I had was thoughtful domain bounding.
147 00:10:55.550 ⇒ 00:10:56.230 Mustafa Raja: which we kind.
148 00:10:56.230 ⇒ 00:11:01.089 Samuel Roberts: kind of have already. You know, like, ABC has that built in.
149 00:11:01.950 ⇒ 00:11:13.830 Mustafa Raja: Yes, I mean, if we are… if we are going, you know, if we are… if we are dividing it into three… three, workflows, that is going to be, zips, then, you know, the RAG, and then the…
150 00:11:13.830 ⇒ 00:11:26.689 Mustafa Raja: cancellation flows that are in the prompt right now. It’s pretty… we know, okay, if it’s a generic question, not for, you know, cancellation flows, go to
151 00:11:26.790 ⇒ 00:11:32.580 Mustafa Raja: the rag, if it is a zip, go to the database one, you know.
152 00:11:32.750 ⇒ 00:11:33.670 Mustafa Raja: Yeah, this…
153 00:11:33.670 ⇒ 00:11:34.050 Samuel Roberts: Yeah.
154 00:11:34.050 ⇒ 00:11:37.219 Mustafa Raja: This does seem, somewhat promising.
155 00:11:37.700 ⇒ 00:11:47.780 Samuel Roberts: Yeah, and this is… this is what had been kind of… I’m glad it pointed it out, because I was… this is what I was sort of thinking when we were moving off of Enid into Mastra, is that we could do something like this.
156 00:11:49.170 ⇒ 00:11:57.260 Samuel Roberts: And then I didn’t want to just make that assumption, I wanted to see some other architectures, but I think this is the best agent… the best way to go.
157 00:11:57.830 ⇒ 00:12:02.650 Samuel Roberts: Exactly what we look like I’m sorry, exactly what the agents look like.
158 00:12:02.980 ⇒ 00:12:13.460 Samuel Roberts: Might be something we have to nail down, because, like, pest knowledge agent, maybe it’s just a generic knowledge agent for all of them, maybe we break them into separate ones.
159 00:12:14.020 ⇒ 00:12:15.799 Samuel Roberts: That’s something we gotta kind of sort out.
160 00:12:17.050 ⇒ 00:12:19.830 Samuel Roberts: But I wanted to get… Go ahead.
161 00:12:20.610 ⇒ 00:12:23.890 Mustafa Raja: So,
162 00:12:24.380 ⇒ 00:12:40.140 Mustafa Raja: So the router that we had initially placed there was for, you know, point to the correct document and look into only that particular document. Are we going to take another stab?
163 00:12:40.410 ⇒ 00:12:42.220 Mustafa Raja: With that, or…
164 00:12:42.510 ⇒ 00:12:45.319 Mustafa Raja: Not because, eventually we turned it off, right?
165 00:12:47.720 ⇒ 00:12:50.869 Samuel Roberts: Yeah, I think this would be a slightly different router.
166 00:12:50.870 ⇒ 00:12:52.390 Mustafa Raja: Because, more based on the…
167 00:12:52.700 ⇒ 00:12:56.369 Samuel Roberts: The context of the question, and Person, right?
168 00:12:56.370 ⇒ 00:12:56.970 Mustafa Raja: Yeah.
169 00:12:57.930 ⇒ 00:13:02.359 Samuel Roberts: correct me if I’m wrong, but when I was looking at the Like, it was… Sorry, go ahead.
170 00:13:03.500 ⇒ 00:13:04.650 Mustafa Raja: Yeah, go on.
171 00:13:04.880 ⇒ 00:13:08.240 Samuel Roberts: I just… I didn’t… I didn’t look at the router that closely, actually, let me…
172 00:13:08.970 ⇒ 00:13:12.770 Mustafa Raja: Yeah, it’s just that it looked into the question.
173 00:13:12.770 ⇒ 00:13:14.010 Samuel Roberts: Dude, look at the… okay.
174 00:13:14.770 ⇒ 00:13:17.250 Samuel Roberts: But it all went to the same agent afterwards, right?
175 00:13:17.520 ⇒ 00:13:26.730 Mustafa Raja: Yeah, it all went to the same agent. The only… the only responsibility the router had was to identify the department. That’s all.
176 00:13:27.070 ⇒ 00:13:27.430 Samuel Roberts: Okay.
177 00:13:28.040 ⇒ 00:13:28.580 Mustafa Raja: Yeah.
178 00:13:28.580 ⇒ 00:13:29.180 Samuel Roberts: So I think we can…
179 00:13:29.180 ⇒ 00:13:36.979 Mustafa Raja: The… the vector store that is attached would have that department’s filter in it.
180 00:13:37.800 ⇒ 00:13:47.400 Mustafa Raja: the vector store. We have removed it, though, but I can give you the query that, or the filter that was there, because I saved it,
181 00:13:47.700 ⇒ 00:13:51.080 Mustafa Raja: Because I didn’t know if, in the future.
182 00:13:51.470 ⇒ 00:13:54.720 Mustafa Raja: we might need it again, right? So, I saved it.
183 00:13:55.140 ⇒ 00:13:58.060 Mustafa Raja: But we’re now moving away from…
184 00:13:58.590 ⇒ 00:14:00.379 Mustafa Raja: And it ends, so it doesn’t matter anymore.
185 00:14:00.400 ⇒ 00:14:02.420 Samuel Roberts: But, that is sort of the…
186 00:14:02.490 ⇒ 00:14:04.750 Mustafa Raja: Architecture that we had.
187 00:14:05.350 ⇒ 00:14:09.889 Samuel Roberts: Okay, so yeah, I think the real question we have now is,
188 00:14:11.350 ⇒ 00:14:14.429 Samuel Roberts: What are the sub… the expert agents we want?
189 00:14:14.710 ⇒ 00:14:16.730 Samuel Roberts: I think, obviously, it’s the zip lookup.
190 00:14:17.970 ⇒ 00:14:19.530 Samuel Roberts: We definitely want…
191 00:14:20.350 ⇒ 00:14:26.890 Samuel Roberts: Well, then the question becomes, do we want to break up the central docs at all by departments or anything?
192 00:14:28.030 ⇒ 00:14:29.370 Mustafa Raja: They are already broken.
193 00:14:30.130 ⇒ 00:14:31.300 Samuel Roberts: I mean by agent.
194 00:14:32.120 ⇒ 00:14:46.089 Mustafa Raja: Oh, yeah, no, no, I don’t think so we would want to break them with the agent, because… So I think… To me, to me, it wouldn’t make sense to break them. We can treat the agent as generic.
195 00:14:46.450 ⇒ 00:14:53.309 Mustafa Raja: generic central dock one, the same way we are doing the client hub thing, you know? We could filter the embeddings.
196 00:14:53.920 ⇒ 00:14:58.939 Samuel Roberts: So could we do… but aren’t we running into problems where we’re seeing the same words in different parts?
197 00:15:00.780 ⇒ 00:15:06.989 Mustafa Raja: I’ll need to look into three others, because I haven’t been with.
198 00:15:06.990 ⇒ 00:15:07.660 Samuel Roberts: Okay.
199 00:15:07.940 ⇒ 00:15:08.640 Samuel Roberts: Yeah.
200 00:15:08.640 ⇒ 00:15:10.549 Mustafa Raja: keeping up with the ABC.
201 00:15:10.550 ⇒ 00:15:16.579 Samuel Roberts: Because my thought maybe is that the RAG agent could be generic, but then we have different tools per…
202 00:15:16.970 ⇒ 00:15:19.379 Samuel Roberts: It knows which tool to call for which…
203 00:15:20.410 ⇒ 00:15:23.050 Samuel Roberts: part of the Docs movie. Or which Doc?
204 00:15:25.110 ⇒ 00:15:33.630 Mustafa Raja: Yeah, so yeah, the table is structured in that way, where we could filter with, departments.
205 00:15:34.170 ⇒ 00:15:34.890 Samuel Roberts: Okay.
206 00:15:35.270 ⇒ 00:15:38.580 Samuel Roberts: Okay, so maybe we don’t need to do it at the tool level, we can just do it as a pass-through.
207 00:15:39.590 ⇒ 00:15:43.180 Mustafa Raja: Yes, we could directly filter at the query level.
208 00:15:43.600 ⇒ 00:15:44.240 Samuel Roberts: Okay.
209 00:15:45.190 ⇒ 00:15:49.620 Samuel Roberts: Okay, so maybe, maybe… I was thinking we.
210 00:15:49.620 ⇒ 00:15:53.379 Mustafa Raja: But do we want to do that? That’s the only question, because.
211 00:15:53.380 ⇒ 00:15:55.090 Samuel Roberts: Yeah, that’s what I wasn’t sure. Maybe it is.
212 00:15:55.090 ⇒ 00:15:59.549 Mustafa Raja: Maybe this is something that I would, I would discuss with Amber.
213 00:15:59.680 ⇒ 00:16:04.630 Mustafa Raja: If you’re having, if you’re facing issues in that regard.
214 00:16:04.750 ⇒ 00:16:06.110 Samuel Roberts: Right.
215 00:16:06.450 ⇒ 00:16:13.899 Mustafa Raja: And if we are, then maybe it’s worth giving a try. We can simultaneously build two versions.
216 00:16:14.640 ⇒ 00:16:15.600 Mustafa Raja: You know?
217 00:16:16.140 ⇒ 00:16:23.410 Samuel Roberts: Yeah, I mean, the nice thing about the monster is that we can have a few of these different agents, and we can just kind of plug them in differently. Yes.
218 00:16:23.720 ⇒ 00:16:24.870 Samuel Roberts: So, okay.
219 00:16:27.270 ⇒ 00:16:28.500 Samuel Roberts: Let me…
220 00:16:29.170 ⇒ 00:16:33.170 Samuel Roberts: Let me think, I just haven’t had a lot of time to really stop and think about this today.
221 00:16:34.490 ⇒ 00:16:35.600 Mustafa Raja: Yeah, no worries.
222 00:16:36.530 ⇒ 00:16:40.619 Samuel Roberts: So, we’re kind of just getting off the dome right now, but I think…
223 00:16:42.180 ⇒ 00:16:47.180 Samuel Roberts: Well, let’s… let’s think about this. Let’s actually… let’s do a Figma, maybe, and start talking about how we could…
224 00:16:47.960 ⇒ 00:16:50.890 Samuel Roberts: Restructure this a little bit.
225 00:16:53.780 ⇒ 00:16:58.139 Samuel Roberts: We got ABC here… So we already…
226 00:17:00.780 ⇒ 00:17:09.460 Samuel Roberts: We already sort of talked about how it’s currently looking, up here… Right.
227 00:17:09.880 ⇒ 00:17:15.119 Samuel Roberts: So now, I think basically the idea is that this routing agent… well, no, I don’t want to map it to this, this is…
228 00:17:16.050 ⇒ 00:17:17.680 Samuel Roberts: I want to kind of think… fresh.
229 00:17:18.609 ⇒ 00:17:23.359 Samuel Roberts: Where… this main Andy logic…
230 00:17:24.700 ⇒ 00:17:26.219 Mustafa Raja: I think we already have this, right?
231 00:17:26.400 ⇒ 00:17:28.520 Mustafa Raja: We, this, this, this piece we are…
232 00:17:28.750 ⇒ 00:17:30.280 Mustafa Raja: The mean and the logic, no?
233 00:17:31.360 ⇒ 00:17:35.029 Samuel Roberts: Well, that’s one thing, I think what I want to try to do is make this more robust, based on what we have.
234 00:17:35.030 ⇒ 00:17:35.770 Mustafa Raja: Yeah.
235 00:17:35.770 ⇒ 00:17:38.409 Samuel Roberts: This is… this is just getting the same…
236 00:17:38.920 ⇒ 00:17:43.319 Samuel Roberts: getting on par with what we have, but I kind of want to see if we can do this multi-agent thing.
237 00:17:44.010 ⇒ 00:17:44.670 Mustafa Raja: Yeah.
238 00:17:45.170 ⇒ 00:17:47.680 Mustafa Raja: So we are further breaking it down, right?
239 00:17:47.990 ⇒ 00:17:48.930 Samuel Roberts: I, well…
240 00:17:50.010 ⇒ 00:17:55.500 Samuel Roberts: I think so. Maybe, maybe it’s unnecessary right now, but I feel like this would solve some of the…
241 00:17:56.790 ⇒ 00:17:59.359 Samuel Roberts: accuracy issues, if we…
242 00:18:00.140 ⇒ 00:18:17.580 Samuel Roberts: further up the chain, figure out… like, the way we had the routing agent to begin with, but making that more robust, and then passing that to maybe two different agents, so that this agent doesn’t have to decide if it’s going superbase or here, but it knows which agent to go to. Maybe that’s overkill, because it’s not as many things as I’m thinking now, but…
243 00:18:17.970 ⇒ 00:18:34.610 Mustafa Raja: You know, part of the reason, routing agent or the router, the router that we had in NITM failed was, CSRs did not have visibility on which department Andy is looking at at a certain point.
244 00:18:34.800 ⇒ 00:18:45.340 Mustafa Raja: And maybe we could add that in response, and if they… if they ask about a department, but Andy’s looking into some other department, they could direct Andy and say.
245 00:18:45.580 ⇒ 00:18:46.190 Samuel Roberts: Yes, I think.
246 00:18:46.190 ⇒ 00:18:46.969 Mustafa Raja: Look into that one.
247 00:18:46.970 ⇒ 00:18:51.549 Samuel Roberts: do is have Andy respond with questions, depending on If there’s not enough.
248 00:18:51.550 ⇒ 00:18:56.469 Mustafa Raja: Yeah, maybe, maybe even if in the answers, it could just say, you know, department this, you know?
249 00:18:57.140 ⇒ 00:18:59.130 Samuel Roberts: Yeah… okay.
250 00:19:05.560 ⇒ 00:19:06.610 Samuel Roberts: Alright.
251 00:19:06.760 ⇒ 00:19:08.459 Mustafa Raja: If you feel good about that.
252 00:19:09.060 ⇒ 00:19:13.960 Samuel Roberts: Yeah, that’s not a bad idea. I’m just trying to… I want to try to… Make sure.
253 00:19:13.960 ⇒ 00:19:14.670 Mustafa Raja: We’re getting somewhere.
254 00:19:14.670 ⇒ 00:19:25.709 Samuel Roberts: with the migration, and, like, that it’s… it’s… it’s clear that this is gonna be better, or more robust, or whatever, but maybe before we start making dramatic changes to the architecture, but…
255 00:19:27.870 ⇒ 00:19:28.960 Samuel Roberts: I don’t know.
256 00:19:30.200 ⇒ 00:19:35.749 Samuel Roberts: I don’t know the triage issues as well as you guys do, so maybe I’m thinking in the wrong,
257 00:19:36.820 ⇒ 00:19:38.780 Samuel Roberts: Down the wrong path, but…
258 00:19:39.710 ⇒ 00:19:44.610 Mustafa Raja: Sorry, yeah, Casey could… Helped us in that regard.
259 00:19:44.980 ⇒ 00:19:45.880 Samuel Roberts: Yeah, that’s why I don’.
260 00:19:45.880 ⇒ 00:19:46.540 Mustafa Raja: And on the road.
261 00:19:46.540 ⇒ 00:19:51.330 Samuel Roberts: go… I know, I don’t wanna… I don’t wanna go too far off a path without…
262 00:19:51.460 ⇒ 00:19:55.000 Samuel Roberts: his input here, because he definitely has more of the domain knowledge here for…
263 00:19:55.000 ⇒ 00:19:55.410 Mustafa Raja: Yeah.
264 00:19:55.410 ⇒ 00:20:01.259 Samuel Roberts: but… Maybe what we can do in the meantime is, let me stop screen sharing real quick.
265 00:20:02.110 ⇒ 00:20:05.849 Samuel Roberts: Let me get up… I don’t think I have the master code up, but let me open…
266 00:20:06.480 ⇒ 00:20:08.990 Samuel Roberts: A right cursor window.
267 00:20:09.360 ⇒ 00:20:17.579 Mustafa Raja: One more thing that I’ve been wanting to ask is, are we good with PG, or do we want to do PG vector?
268 00:20:17.840 ⇒ 00:20:22.500 Mustafa Raja: Because that is something that is stuck with,
269 00:20:22.740 ⇒ 00:20:28.820 Mustafa Raja: to the generic client hubs, and then we have the same logic over here. I just want.
270 00:20:28.820 ⇒ 00:20:29.460 Samuel Roberts: You’re racial…
271 00:20:29.460 ⇒ 00:20:31.620 Mustafa Raja: If this is production ready or not.
272 00:20:32.440 ⇒ 00:20:37.149 Samuel Roberts: Yeah, let me actually… I haven’t been able to, like I said, test this as much, so let me…
273 00:20:37.150 ⇒ 00:20:37.660 Mustafa Raja: Yeah.
274 00:20:37.660 ⇒ 00:20:43.099 Samuel Roberts: Let me get… get in here real quick and see. And maybe you can walk me through exactly where the…
275 00:20:43.100 ⇒ 00:20:49.470 Mustafa Raja: Yeah, yeah. So the master… let’s get into master, one, because it’s a lot simpler.
276 00:20:49.610 ⇒ 00:20:56.580 Mustafa Raja: Sorry, master one, I meant ABC1, because the code base is a lot, a lot, a lot slim.
277 00:20:56.580 ⇒ 00:20:57.180 Samuel Roberts: Yeah.
278 00:20:57.670 ⇒ 00:20:59.149 Mustafa Raja: So, do you want me to share?
279 00:20:59.150 ⇒ 00:21:00.810 Samuel Roberts: Sure. Yeah, perfect.
280 00:21:01.990 ⇒ 00:21:04.930 Mustafa Raja: Okay, so… so here we have the tool.
281 00:21:05.090 ⇒ 00:21:15.079 Mustafa Raja: This, this isn’t the tool, this is. So, we take in, yeah, this isn’t that, okay.
282 00:21:15.390 ⇒ 00:21:18.129 Mustafa Raja: Yeah, so input schema, we take in the query, right?
283 00:21:18.270 ⇒ 00:21:40.959 Mustafa Raja: So the query, whatever, the agent passes in, to search for in the vector table, right? And then whatever the output schema is, and then what we do is just go ahead and, you know, search the vector. And then here, we just, you know, this client is, coming from PG.
284 00:21:41.170 ⇒ 00:21:52.649 Mustafa Raja: reconnect, and this connects every time, which… and this is exactly the point where the issue was, you know, I was having the issue.
285 00:21:52.650 ⇒ 00:21:53.110 Samuel Roberts: So this…
286 00:21:53.110 ⇒ 00:22:06.019 Mustafa Raja: client disconnects every time, but with PGVector, what used to happen is it would sometimes… it would, you know, connect, and a lot of the times it would not connect.
287 00:22:06.020 ⇒ 00:22:21.779 Mustafa Raja: And that was frustrating. I even tried changing networks, my Wi-Fi to my, you know, mobile data, you know, because I felt that it just… it could just be my internet acting up, right?
288 00:22:23.810 ⇒ 00:22:38.989 Mustafa Raja: So maybe this is something that either Casey could, you know… we just… what we need to do is we just need to plug in the PG vector here, and this would change a little bit. The rest of the stuff is going to be exactly the same, you know.
289 00:22:39.430 ⇒ 00:22:40.810 Mustafa Raja: If that makes sense.
290 00:22:41.590 ⇒ 00:22:43.469 Samuel Roberts: Yeah, I’m just trying to think…
291 00:22:43.470 ⇒ 00:23:08.290 Mustafa Raja: Anyways, let me, let me, let me, let me go forward, let me complete, and then, what we do is, yeah, we just connect, and then we just run the query. So, here we have the vector, which we generate over here, so, so whatever the query, that, that’s, that’s been sent by our, what’s it called, agent, it’s converted into embedd
292 00:23:08.290 ⇒ 00:23:15.020 Mustafa Raja: ratings, and then… and then this is just a simple match. And then, one more thing is, this is exactly the same response.
293 00:23:15.020 ⇒ 00:23:18.519 Mustafa Raja: Andy would do in any tent. So, we know that this…
294 00:23:18.560 ⇒ 00:23:21.039 Mustafa Raja: this stuff is pretty good. And…
295 00:23:21.040 ⇒ 00:23:21.360 Samuel Roberts: So…
296 00:23:21.360 ⇒ 00:23:35.539 Mustafa Raja: This is, this is our response coming in. We have, so we are limit… limiting to 10, 10 results per query, so this is what it is. Okay. So, match rate, not so, not so confident, but…
297 00:23:35.820 ⇒ 00:23:40.120 Mustafa Raja: Yeah, we would have a similar situation in Enton.
298 00:23:40.590 ⇒ 00:23:44.360 Samuel Roberts: That is good, though, that we can really see it here a little bit better, too, while we’re testing.
299 00:23:44.640 ⇒ 00:23:50.099 Samuel Roberts: Yeah, oh, one more thing. We could actually, you know, just go ahead and, you know, do this. Yeah, I was gonna say, this is…
300 00:23:50.100 ⇒ 00:23:50.889 Mustafa Raja: whole flow.
301 00:23:50.890 ⇒ 00:23:52.570 Samuel Roberts: Exactly, exactly.
302 00:23:52.570 ⇒ 00:24:05.190 Mustafa Raja: Yeah, the other day, when I was testing this query tool, I tested it from here and not the whole agent, because that’s longer flow, and I just want to test the tool, and this is a pretty good, alternative. I like it.
303 00:24:05.190 ⇒ 00:24:09.580 Samuel Roberts: That’s great, yeah. This is why I was really excited about using this flow here.
304 00:24:09.580 ⇒ 00:24:15.910 Mustafa Raja: Yeah, and also, if you notice, this is in… what’s it called? Cursor.
305 00:24:15.910 ⇒ 00:24:16.480 Samuel Roberts: Oh, I know.
306 00:24:16.480 ⇒ 00:24:17.190 Mustafa Raja: Yeah, surprising.
307 00:24:17.190 ⇒ 00:24:18.879 Samuel Roberts: Cursor’s been so good, yeah.
308 00:24:19.080 ⇒ 00:24:23.400 Mustafa Raja: Yeah, and for UI, it’s… it’s crazy good, because
309 00:24:23.570 ⇒ 00:24:28.420 Mustafa Raja: Just select this, and it… it just goes into the chat. Yeah.
310 00:24:28.420 ⇒ 00:24:32.560 Samuel Roberts: Oh, I didn’t even realize you could do that much! Oh my goodness, that’s gonna change my run-in.
311 00:24:32.560 ⇒ 00:24:40.740 Mustafa Raja: Yeah, maybe this is something that we would… we would, you know, love in Lilo or something, because I believe we would be doing something with the UI, right?
312 00:24:40.990 ⇒ 00:24:44.000 Samuel Roberts: Yeah, there’ll definitely be a lot of UI stuff. I’ll make sure…
313 00:24:44.390 ⇒ 00:24:46.260 Samuel Roberts: Yeah, that’s… yes, okay, that’s great.
314 00:24:46.460 ⇒ 00:24:48.219 Samuel Roberts: Oh my god, I didn’t realize it had that much.
315 00:24:48.220 ⇒ 00:24:49.369 Mustafa Raja: That’s all good.
316 00:24:49.810 ⇒ 00:24:54.420 Mustafa Raja: Yeah. So, yeah, so this is the gist of it, you know?
317 00:24:54.890 ⇒ 00:25:00.420 Samuel Roberts: Okay, stepping back a second, for the PGPG vector thing,
318 00:25:00.620 ⇒ 00:25:04.529 Samuel Roberts: how much different does that query look if you’re using PGVector?
319 00:25:04.930 ⇒ 00:25:12.689 Mustafa Raja: So PGVector, we don’t have to create the query. PGVector handles the query, we just have to pass in the embeddings, you know?
320 00:25:12.690 ⇒ 00:25:14.580 Samuel Roberts: Okay. So, I honestly think what…
321 00:25:14.580 ⇒ 00:25:16.719 Mustafa Raja: There is… it’s just an apple.
322 00:25:17.200 ⇒ 00:25:24.399 Samuel Roberts: Exactly, I was gonna say the same thing. I think we’re fine like this, because the PGVector extension is what’s doing the work here.
323 00:25:25.190 ⇒ 00:25:28.270 Samuel Roberts: And that’s in the Postgres instance in Subabase.
324 00:25:28.570 ⇒ 00:25:29.510 Mustafa Raja: Yes. Right.
325 00:25:30.060 ⇒ 00:25:33.500 Samuel Roberts: So I think this is just a wrapper… that…
326 00:25:33.720 ⇒ 00:25:38.259 Samuel Roberts: makes it easier to interact with that than doing it manually with the SQL.
327 00:25:40.970 ⇒ 00:25:52.079 Mustafa Raja: Yeah, yeah. I believe. So, it’s just… what I believe… what I believe is it’s just a wrapper on… on a query like this or something, and then, yeah, that’s pretty much it.
328 00:25:52.330 ⇒ 00:25:56.259 Samuel Roberts: Yeah, I don’t think there’s a reason to worry about it if we’re having issues.
329 00:25:56.260 ⇒ 00:26:03.779 Mustafa Raja: But I would still want to, you know, have a second eye, I don’t know. Yeah.
330 00:26:05.340 ⇒ 00:26:08.579 Samuel Roberts: No, I understand. I will get in there and do some testing.
331 00:26:09.270 ⇒ 00:26:21.860 Mustafa Raja: Okay, yeah, that would be really helpful, but this is the whole architecture, and this is exactly the same thing that is… that I have implemented for Slack and Zoom tools in our platform, you know?
332 00:26:22.240 ⇒ 00:26:24.709 Mustafa Raja: For generic, client hubs.
333 00:26:25.160 ⇒ 00:26:26.980 Samuel Roberts: Yeah, I mean, I think for, like…
334 00:26:27.580 ⇒ 00:26:33.359 Samuel Roberts: for these tools, to just, you know, do the vector query, that makes total sense. I think the question becomes.
335 00:26:33.870 ⇒ 00:26:39.150 Samuel Roberts: And maybe this is where we make different tools that look specifically at certain parts.
336 00:26:39.780 ⇒ 00:26:44.080 Samuel Roberts: Where, like, this is a vector search, but you said there’s a…
337 00:26:44.080 ⇒ 00:26:44.510 Mustafa Raja: I mean.
338 00:26:44.510 ⇒ 00:26:47.040 Samuel Roberts: We store what documents and stuff everything is?
339 00:26:47.400 ⇒ 00:26:51.320 Mustafa Raja: Yes, let me actually pull the database, too.
340 00:26:52.160 ⇒ 00:27:01.369 Mustafa Raja: So, the database is structured in a way where this… this query can filter anything. If you want any specifics.
341 00:27:02.050 ⇒ 00:27:08.860 Mustafa Raja: So, I don’t think we need to, you know, break down on there if we can handle stuff with filters.
342 00:27:08.970 ⇒ 00:27:12.170 Mustafa Raja: But, yeah, let me… let me pull up.
343 00:27:13.260 ⇒ 00:27:18.989 Samuel Roberts: Because even… even if we filter with different agents, if the… if the… or different tools, I should say.
344 00:27:19.170 ⇒ 00:27:23.490 Samuel Roberts: Like, then maybe they’re doing very similar things, but we’ve…
345 00:27:24.010 ⇒ 00:27:33.349 Samuel Roberts: you know, we’ve simplified it for the LLM to know, okay, I just call this tool, and it searches this document, and call this tool, it searches that… yes, perfect, okay.
346 00:27:33.570 ⇒ 00:27:35.359 Samuel Roberts: And that’s in the metadata?
347 00:27:35.760 ⇒ 00:27:36.500 Mustafa Raja: Yes.
348 00:27:37.060 ⇒ 00:27:37.730 Samuel Roberts: Okay.
349 00:27:38.720 ⇒ 00:27:49.069 Mustafa Raja: So I did… I did this in metadata, because I, prior to this, I worked in, Llama Index and Langchen, and they… they… they highly suggest that.
350 00:27:49.170 ⇒ 00:27:59.769 Mustafa Raja: I don’t think Master suggests this, but this still isn’t a limitation to us, because we can… it’s just Postgres, and we can just save.
351 00:27:59.770 ⇒ 00:28:03.000 Samuel Roberts: Yeah, and it’s… it’s JSONB, it’s JSON, so we’re.
352 00:28:03.000 ⇒ 00:28:03.760 Mustafa Raja: Yeah.
353 00:28:04.010 ⇒ 00:28:09.959 Mustafa Raja: Yeah, so we… I believe we are still good. We… we… we… we can…
354 00:28:10.460 ⇒ 00:28:19.780 Mustafa Raja: Specify departments, if we can detect, at the starting point which department is the question intended for.
355 00:28:20.570 ⇒ 00:28:22.010 Samuel Roberts: Yes. Okay.
356 00:28:25.140 ⇒ 00:28:25.970 Samuel Roberts: Okay.
357 00:28:26.100 ⇒ 00:28:26.920 Samuel Roberts: No.
358 00:28:27.620 ⇒ 00:28:28.140 Samuel Roberts: Okay.
359 00:28:28.140 ⇒ 00:28:40.870 Mustafa Raja: And, part of the limitation that, Aniton had is that maybe a question could be for two departments at the same time, and this is something that Aniton cannot handle. So, Aniton’s.
360 00:28:40.870 ⇒ 00:28:41.380 Samuel Roberts: Great.
361 00:28:41.380 ⇒ 00:28:45.020 Mustafa Raja: Printer is an AND filter, you know?
362 00:28:45.020 ⇒ 00:28:47.349 Samuel Roberts: Right, I forgot we had that problem, yeah, yeah, yeah.
363 00:28:47.860 ⇒ 00:28:55.719 Mustafa Raja: So… so now, now here, here we… we have both of the filters because everything is in our control.
364 00:28:56.390 ⇒ 00:29:00.700 Samuel Roberts: Yes, okay, that’s good. I forgot that they only had the AND filter, that’s crazy.
365 00:29:01.780 ⇒ 00:29:06.389 Mustafa Raja: Yeah, that’s a crazy limitation to have for such a big… Platform, to be honest.
366 00:29:06.390 ⇒ 00:29:07.730 Samuel Roberts: Isn’t it, isn’t it…
367 00:29:07.730 ⇒ 00:29:08.580 Mustafa Raja: Yeah.
368 00:29:09.650 ⇒ 00:29:13.069 Mustafa Raja: It limits a lot of the possibilities, you know?
369 00:29:13.560 ⇒ 00:29:14.190 Samuel Roberts: Yeah.
370 00:29:15.040 ⇒ 00:29:18.759 Mustafa Raja: For that reason, I had to… I had to include this key.
371 00:29:20.190 ⇒ 00:29:21.490 Samuel Roberts: Oh my goodness, okay.
372 00:29:22.530 ⇒ 00:29:24.950 Samuel Roberts: It’s such an honor. Oh my gosh, yeah.
373 00:29:25.070 ⇒ 00:29:42.050 Mustafa Raja: Because we cannot even remove the filter. So, the reason I added this, if the routing agent that decides the department is not sure, and wants to run the query.
374 00:29:42.170 ⇒ 00:29:49.319 Mustafa Raja: We cannot do this because the value should be a valid value, and we cannot say, okay, if
375 00:29:49.890 ⇒ 00:29:53.579 Mustafa Raja: We cannot remove the filter on runtime.
376 00:29:55.760 ⇒ 00:29:58.150 Samuel Roberts: Okay, so… Now we…
377 00:29:58.150 ⇒ 00:30:00.909 Mustafa Raja: So that is why this existed, yeah.
378 00:30:00.910 ⇒ 00:30:08.569 Samuel Roberts: Okay, so here’s the other thing, though. Was Nadet able to respond and ask a question of the CSR in the middle?
379 00:30:09.300 ⇒ 00:30:10.010 Mustafa Raja: Yes!
380 00:30:11.590 ⇒ 00:30:12.000 Samuel Roberts: So, was it.
381 00:30:12.000 ⇒ 00:30:22.130 Mustafa Raja: But it did not do a good, good job, but part of the reason, but I believe part of the reason was the system prompt is too bloated.
382 00:30:22.660 ⇒ 00:30:27.499 Samuel Roberts: Yes, okay, so that’s two… so that’s two things we can do here. One, I think we should…
383 00:30:28.400 ⇒ 00:30:35.750 Samuel Roberts: Take, like, strip down that system prompt and put all that information that we were using into another tool.
384 00:30:36.840 ⇒ 00:30:37.500 Mustafa Raja: 3 cheers.
385 00:30:37.500 ⇒ 00:30:37.909 Samuel Roberts: So, like.
386 00:30:37.910 ⇒ 00:30:38.440 Mustafa Raja: again?
387 00:30:39.050 ⇒ 00:30:42.640 Samuel Roberts: All the… what is it, the cancellation flows, or whatever we were putting into the…
388 00:30:43.360 ⇒ 00:30:48.039 Mustafa Raja: Let’s… let me… let me… let me go to Elang Fuse, and let’s…
389 00:30:48.040 ⇒ 00:30:49.180 Samuel Roberts: Yeah, yeah, totally.
390 00:30:49.180 ⇒ 00:30:52.680 Mustafa Raja: Let’s see whatever… whatever their holding was.
391 00:30:56.890 ⇒ 00:31:07.699 Mustafa Raja: Okay, and then bronze… And, okay, so… Values, queues…
392 00:31:09.630 ⇒ 00:31:14.679 Mustafa Raja: Yeah, I think the prompt ends here. And now let’s see how long this is.
393 00:31:16.340 ⇒ 00:31:18.720 Samuel Roberts: It’s, like, the vast majority of it, look at that.
394 00:31:19.370 ⇒ 00:31:20.250 Mustafa Raja: Yeah.
395 00:31:20.550 ⇒ 00:31:23.230 Mustafa Raja: Yeah, it just goes and goes and goes.
396 00:31:23.230 ⇒ 00:31:29.050 Samuel Roberts: Okay, yeah, so what we can do now is… Either we embed all that.
397 00:31:29.250 ⇒ 00:31:36.660 Samuel Roberts: Or we make a tool that just adds it as more context, which is probably all we need to do for now, at least to get it.
398 00:31:36.660 ⇒ 00:31:37.729 Mustafa Raja: Say that again?
399 00:31:38.390 ⇒ 00:31:40.569 Samuel Roberts: So how often was this changing?
400 00:31:40.690 ⇒ 00:31:48.220 Mustafa Raja: This, I think this… these things were added by Casey,
401 00:31:49.690 ⇒ 00:31:53.079 Mustafa Raja: So, I would not have context on this.
402 00:31:53.080 ⇒ 00:31:59.699 Samuel Roberts: Okay, so let’s assume that this stuff is pretty stable, and if not, we can change it later, but I think the idea here is that
403 00:31:59.890 ⇒ 00:32:06.729 Samuel Roberts: Instead of, like, the main part will be the main agent prompt, and when it thinks it needs to look up
404 00:32:07.950 ⇒ 00:32:14.249 Samuel Roberts: Cancellation flows, or whatever other stuff is in that context, it’ll call another tool that will provide that.
405 00:32:17.190 ⇒ 00:32:21.089 Mustafa Raja: Sufficient attributes… let me search it up when it ends.
406 00:32:21.840 ⇒ 00:32:24.869 Samuel Roberts: So, I think, yeah, there’s that, there’s cancellation flow…
407 00:32:24.870 ⇒ 00:32:26.340 Mustafa Raja: Transmission flow.
408 00:32:27.930 ⇒ 00:32:33.470 Mustafa Raja: Okay, when this ends… This ends, yeah, at the bottom only.
409 00:32:33.470 ⇒ 00:32:34.140 Samuel Roberts: Okay.
410 00:32:34.380 ⇒ 00:32:38.629 Samuel Roberts: So, basically, I think what we should do is we’ll have… okay, let me just…
411 00:32:39.000 ⇒ 00:32:46.260 Samuel Roberts: We’ll have, like, the routing agent, maybe, that will figure out departments and stuff, or figure out if it’s asking about cancellations or whatever.
412 00:32:46.540 ⇒ 00:32:53.990 Samuel Roberts: That will route to… The main agent here… Or… A cancellation agent.
413 00:32:55.800 ⇒ 00:32:59.259 Samuel Roberts: The cancellation agent will be the one that has more stuff.
414 00:32:59.760 ⇒ 00:33:01.729 Samuel Roberts: But the main agent doesn’t need all that.
415 00:33:03.120 ⇒ 00:33:03.850 Mustafa Raja: Okay.
416 00:33:04.840 ⇒ 00:33:11.069 Mustafa Raja: Maybe what we, let me see when Casey’s back on…
417 00:33:11.700 ⇒ 00:33:14.519 Samuel Roberts: I think he’s out till Wednesday, but we can maybe message him and get a question.
418 00:33:14.520 ⇒ 00:33:15.300 Mustafa Raja: Oh.
419 00:33:16.220 ⇒ 00:33:19.009 Mustafa Raja: Yeah, I was just wondering if,
420 00:33:20.120 ⇒ 00:33:25.700 Mustafa Raja: If he has… so, dropping or cancellation of attributes, then there’s cancellation flow.
421 00:33:26.010 ⇒ 00:33:37.249 Mustafa Raja: And then there is estimator quotes. Maybe, yeah, yeah, the router agent can have context on these three things, right? If the questions are about these, route to that.
422 00:33:37.940 ⇒ 00:33:43.999 Samuel Roberts: Yeah, exactly, and then that’s basically stripping it out of the main heavy lifting of the rag.
423 00:33:44.370 ⇒ 00:33:50.280 Samuel Roberts: For the central dock, and the rag… or the lookup rag for the… Zips.
424 00:33:52.070 ⇒ 00:33:52.690 Samuel Roberts: So…
425 00:33:53.170 ⇒ 00:33:56.700 Mustafa Raja: And then we can also trim down a lot of this stuff, because…
426 00:33:57.160 ⇒ 00:34:00.800 Mustafa Raja: Some of this is relevant also.
427 00:34:01.270 ⇒ 00:34:02.370 Samuel Roberts: Yeah, but we can, we can…
428 00:34:02.370 ⇒ 00:34:04.670 Mustafa Raja: like.
429 00:34:05.660 ⇒ 00:34:06.590 Samuel Roberts: A tool you use?
430 00:34:06.590 ⇒ 00:34:07.890 Mustafa Raja: does not exist.
431 00:34:08.469 ⇒ 00:34:13.729 Samuel Roberts: Yeah, exactly. Okay, so we can definitely clean a lot of this up, which is probably gonna help anyway, even if it were still an end.
432 00:34:14.469 ⇒ 00:34:16.569 Mustafa Raja: And this says non-exist.
433 00:34:17.489 ⇒ 00:34:18.119 Samuel Roberts: Okay.
434 00:34:22.960 ⇒ 00:34:33.900 Mustafa Raja: I don’t think we need memory instructions, to be honest, because that is something that would be maintained manually, right?
435 00:34:34.170 ⇒ 00:34:38.880 Samuel Roberts: It… it is, but I think it’s good for… to know that we don’t want to use that yet.
436 00:34:39.630 ⇒ 00:34:42.260 Samuel Roberts: I don’t know, I wouldn’t change all that yet. Here’s my thought.
437 00:34:42.260 ⇒ 00:34:42.610 Mustafa Raja: Okay.
438 00:34:42.610 ⇒ 00:34:45.160 Samuel Roberts: for next steps, I would say.
439 00:34:45.659 ⇒ 00:34:52.969 Samuel Roberts: Strip out those parts of the context, and make separate agents with that in their context as well.
440 00:34:53.340 ⇒ 00:34:54.020 Mustafa Raja: Okay.
441 00:34:54.630 ⇒ 00:34:55.310 Mustafa Raja: Okay.
442 00:34:55.310 ⇒ 00:34:57.709 Samuel Roberts: Either that or a tool call.
443 00:34:59.089 ⇒ 00:34:59.779 Samuel Roberts: I don’t know.
444 00:35:01.740 ⇒ 00:35:05.250 Mustafa Raja: Dull guy, mmm…
445 00:35:06.090 ⇒ 00:35:07.220 Samuel Roberts: I think, I think an agent makes sense.
446 00:35:07.220 ⇒ 00:35:10.720 Mustafa Raja: Dool call is also possible, because this isn’t…
447 00:35:10.720 ⇒ 00:35:11.120 Samuel Roberts: Yeah.
448 00:35:11.120 ⇒ 00:35:16.709 Mustafa Raja: This isn’t even needed… this does not even need to be, you know, embedded also, right?
449 00:35:17.140 ⇒ 00:35:19.349 Samuel Roberts: Right, it’s a So we can pass in.
450 00:35:19.350 ⇒ 00:35:19.690 Mustafa Raja: Yeah.
451 00:35:19.690 ⇒ 00:35:20.220 Samuel Roberts: Yeah.
452 00:35:20.490 ⇒ 00:35:24.189 Mustafa Raja: Yeah, this could just live as is, because it’s already living as is, right?
453 00:35:24.700 ⇒ 00:35:32.079 Samuel Roberts: Right, right. So, yeah, I think all we need to do is make sure that we just don’t include it every time, we only include it when the questions are pertinent.
454 00:35:33.030 ⇒ 00:35:37.910 Samuel Roberts: And so maybe there’s a cancellation tool that gets the info for that and passes it into the agent.
455 00:35:38.030 ⇒ 00:35:41.320 Samuel Roberts: and a… Whatever else.
456 00:35:41.720 ⇒ 00:35:43.190 Samuel Roberts: what other folks? So I would say…
457 00:35:43.190 ⇒ 00:35:43.990 Mustafa Raja: And I…
458 00:35:43.990 ⇒ 00:35:45.110 Samuel Roberts: Yeah, perfect.
459 00:35:45.110 ⇒ 00:35:55.610 Mustafa Raja: I’ll do that, and I’ll ask Casey, if he has, you know, sample questions of, that, that CSRs would ask that needs…
460 00:35:55.610 ⇒ 00:35:56.010 Samuel Roberts: S.
461 00:35:56.010 ⇒ 00:36:07.309 Mustafa Raja: information from these ones, so we know, you know, so router has context that, okay, if this sort of stuff is being asked, then this belongs to the, you know.
462 00:36:07.630 ⇒ 00:36:10.090 Samuel Roberts: Yep, you’re, you’re, yeah, you’re right on track there.
463 00:36:11.010 ⇒ 00:36:14.469 Mustafa Raja: Okay, so, so for the further steps…
464 00:36:14.770 ⇒ 00:36:22.440 Mustafa Raja: And I’ll try to push something today. So the further step is… sorry, the further steps,
465 00:36:23.080 ⇒ 00:36:26.460 Mustafa Raja: Router, and then strip out the main agent into three, right?
466 00:36:27.660 ⇒ 00:36:31.010 Samuel Roberts: Yeah, or however many are there, I didn’t… I didn’t see how many there were there, but…
467 00:36:31.390 ⇒ 00:36:38.790 Mustafa Raja: So, by 3, I mean the database. So, by database, I mean, SQL to…
468 00:36:39.110 ⇒ 00:36:39.789 Samuel Roberts: Oh, yes, yes.
469 00:36:40.210 ⇒ 00:36:40.609 Samuel Roberts: In that case.
470 00:36:40.610 ⇒ 00:36:49.150 Mustafa Raja: That, and then, what’s it called? The… Central docs, and then this… this fixed stuff.
471 00:36:49.490 ⇒ 00:37:06.580 Mustafa Raja: And does this also include… does this also mean that, sorry, do I need to also ask, CSRs if they’re asking questions, in Central Docs, for which department do they need the answer from?
472 00:37:07.740 ⇒ 00:37:15.109 Samuel Roberts: Possibly… I think if there’s a way… so that routing agent was doing a lookup initially, right?
473 00:37:16.010 ⇒ 00:37:26.530 Mustafa Raja: Yes, but that got all… that got confusing, because the CSRs that were, you know, testing it were in a lot of departments.
474 00:37:26.530 ⇒ 00:37:27.020 Samuel Roberts: That’s what I was.
475 00:37:27.020 ⇒ 00:37:27.860 Mustafa Raja: So there’s a mile.
476 00:37:27.860 ⇒ 00:37:28.600 Samuel Roberts: But yeah.
477 00:37:29.210 ⇒ 00:37:45.200 Mustafa Raja: My logic initially was, okay, a CSR is, you know, associated with one, at max, two departments, and this is a big fuss, so this is… this should just be okay. But turns out, some of this, some of these, power users
478 00:37:45.200 ⇒ 00:37:55.840 Mustafa Raja: Really were associated, 4, sometimes, sometimes 3 departments, and yeah, yeah, at that point, that was just useless.
479 00:37:56.020 ⇒ 00:37:56.820 Mustafa Raja: the sheet.
480 00:37:56.820 ⇒ 00:38:06.210 Samuel Roberts: Well, that’s still… still interesting to potentially do, because we could have the agent know what department they’re in, and if they’re in multiple, it could ask for clarification. If.
481 00:38:06.210 ⇒ 00:38:11.019 Mustafa Raja: That is exactly the flow that I had, you know?
482 00:38:11.020 ⇒ 00:38:21.599 Samuel Roberts: Yeah, exactly. But I think… I think what we could do more interestingly is once it does the lookup, if there are multiple departments, it could also ask for clarification then, before it responds.
483 00:38:22.350 ⇒ 00:38:27.380 Mustafa Raja: Yeah, yeah, this is the thing that was implemented in Edmonton, but I think…
484 00:38:27.640 ⇒ 00:38:32.660 Samuel Roberts: Oh, was it? Okay, I thought it was just the… which departments they were in, but if the response has multiple departments…
485 00:38:33.060 ⇒ 00:38:48.600 Mustafa Raja: Yeah, so, yeah, yeah, so… No, no, no, oh, so you mean if the final response by the, by the, after the router agent, when the main agent would work, if it finds contacts in multiple departments?
486 00:38:48.900 ⇒ 00:38:49.940 Mustafa Raja: Then we should knock it.
487 00:38:49.940 ⇒ 00:38:52.839 Samuel Roberts: Does… potentially, yeah, I think that’s not a bad…
488 00:38:54.280 ⇒ 00:39:06.389 Mustafa Raja: To be honest, we are at a point where we could, just filter multiple departments, right, in, what’s it called?
489 00:39:06.650 ⇒ 00:39:09.940 Mustafa Raja: Central dogs, you know, if that makes sense.
490 00:39:09.940 ⇒ 00:39:10.500 Samuel Roberts: Yeah.
491 00:39:11.120 ⇒ 00:39:17.690 Samuel Roberts: Yeah, that’s what I was wondering, if we want to say, like, instead of having one tool that’s generic.
492 00:39:18.720 ⇒ 00:39:24.280 Samuel Roberts: And then the agent has to know which ones it can take as an… no, actually, that’s the way to do it, right there. How many departments do we have?
493 00:39:25.270 ⇒ 00:39:25.880 Samuel Roberts: That are in the.
494 00:39:25.880 ⇒ 00:39:36.760 Mustafa Raja: I think 5 or 6, I think it’s five departments, and then one coverage dock. The coverage dock just has a high-level view of all departments.
495 00:39:36.940 ⇒ 00:39:42.980 Samuel Roberts: Yeah, so I would say… For the vector lookup, we could maybe have it specify the department?
496 00:39:44.040 ⇒ 00:39:44.860 Samuel Roberts: As an option.
497 00:39:44.860 ⇒ 00:39:51.010 Mustafa Raja: Your lookup should specify the department, as in we are filtering for that particular department only.
498 00:39:51.010 ⇒ 00:39:57.709 Samuel Roberts: And then I think, and I hope, that if Monster doesn’t know that input, it will ask for it on its own.
499 00:40:00.050 ⇒ 00:40:03.110 Mustafa Raja: Okay, so if I, okay.
500 00:40:03.970 ⇒ 00:40:04.449 Samuel Roberts: Does that make sense?
501 00:40:04.450 ⇒ 00:40:05.459 Mustafa Raja: And .
502 00:40:05.460 ⇒ 00:40:21.910 Samuel Roberts: I can’t make a tool call if it doesn’t know the department, so it might say, oh, I need to… so it might do that, but we’ll have to see how the agent works out. So I say play around with that and see how that works as well, and that will dictate whether or not we need to be more, deterministic with it, or if we can let the agent determine when it needs to ask.
503 00:40:22.340 ⇒ 00:40:28.469 Mustafa Raja: And we could explicitly ask it to, you know, do not assume. You have to be sure.
504 00:40:28.470 ⇒ 00:40:36.450 Samuel Roberts: Definitely put that in there, yeah, do not assume, and maybe at some point we can add back that logic to look up the user and stuff, but for now, just say do not assume department. Yeah.
505 00:40:36.450 ⇒ 00:40:50.180 Mustafa Raja: So, does this not mean that, at least for a session, it will ask, at least once from the CSR which department are they looking for, right?
506 00:40:51.660 ⇒ 00:40:55.129 Samuel Roberts: Yeah, so I actually… that’s a good question. How are these sessions being set up?
507 00:40:55.260 ⇒ 00:40:59.150 Samuel Roberts: Because the session could have multiple… is it… Yeah.
508 00:40:59.150 ⇒ 00:41:00.700 Mustafa Raja: Oh, yeah.
509 00:41:00.900 ⇒ 00:41:04.609 Samuel Roberts: Yeah, well, okay. Every session is not a different customer interaction.
510 00:41:04.610 ⇒ 00:41:06.169 Mustafa Raja: Okay, yeah, yeah, yeah, yeah.
511 00:41:06.170 ⇒ 00:41:07.780 Samuel Roberts: So I might have to ask every time.
512 00:41:07.970 ⇒ 00:41:09.750 Samuel Roberts: And not, not assume, ever.
513 00:41:09.750 ⇒ 00:41:11.520 Mustafa Raja: Okay. Unless…
514 00:41:11.520 ⇒ 00:41:15.950 Samuel Roberts: Unless we want to add that user lookup, CSR lookup, but…
515 00:41:16.640 ⇒ 00:41:20.950 Mustafa Raja: Okay, okay. At that point, do we really need memory?
516 00:41:22.750 ⇒ 00:41:26.389 Samuel Roberts: Well, we want some memory because we want them to be able to chat with it in an individual…
517 00:41:26.900 ⇒ 00:41:28.369 Mustafa Raja: Oh yeah, yeah, you’re right.
518 00:41:28.370 ⇒ 00:41:38.759 Samuel Roberts: individual time, but maybe they have to open a new one every time? But that just seems tedious with the ChiChat setup, but… I don’t know. I wouldn’t stress about that too much yet, because we can get to that bridge…
519 00:41:38.960 ⇒ 00:41:39.890 Samuel Roberts: Soon.
520 00:41:40.080 ⇒ 00:41:41.739 Samuel Roberts: But just for now.
521 00:41:42.660 ⇒ 00:41:52.670 Samuel Roberts: let’s work on what we just said, where we’ll separate the context out into separate agents, or tools, whatever works best, maybe. Try it both ways, I suppose.
522 00:41:52.800 ⇒ 00:42:02.670 Samuel Roberts: And then… The… try out that vector lookup with department, And maybe…
523 00:42:04.180 ⇒ 00:42:11.519 Samuel Roberts: Well, okay, depending on how long that all takes, I would say that the two things that I want to kind of understand for the department side is either
524 00:42:12.280 ⇒ 00:42:14.260 Samuel Roberts: Clarifying the department up front.
525 00:42:15.960 ⇒ 00:42:24.370 Samuel Roberts: Or doing the lookup, seeing multiple departments, and then clarifying, and maybe we can get more… a higher level of,
526 00:42:24.860 ⇒ 00:42:27.810 Samuel Roberts: Probability that it’s right.
527 00:42:29.840 ⇒ 00:42:30.170 Samuel Roberts: I would say.
528 00:42:30.170 ⇒ 00:42:41.390 Mustafa Raja: Yeah, you’re right. So what would then happen is we would narrow down to the department, and then if the answer is wrong, we would know there’s something wrong with the central dog.
529 00:42:42.080 ⇒ 00:42:43.929 Samuel Roberts: Yeah, so I would say try that out and see.
530 00:42:44.600 ⇒ 00:42:45.909 Samuel Roberts: Which way works better?
531 00:42:46.420 ⇒ 00:42:47.430 Mustafa Raja: Yeah, that, that…
532 00:42:47.430 ⇒ 00:42:58.180 Samuel Roberts: Either before or after, because depending on the question, it might be very clear, and depending on the question, it might be ambiguous. And we kind of want to handle both of those cases pretty seamlessly.
533 00:42:58.900 ⇒ 00:43:00.030 Mustafa Raja: Okay.
534 00:43:01.200 ⇒ 00:43:08.069 Mustafa Raja: So, do we have a deadline or timeline for this, migration? Because I, I don’t have context on that, sorry.
535 00:43:08.420 ⇒ 00:43:23.560 Samuel Roberts: We’re getting there. I wanted… I kind of wanted to make sure that some of this stuff was working as we expected before I started saying we’re gonna start moving everything, and that’s kind of where we’re still at right now. So I’m hoping to put that together, like.
536 00:43:24.060 ⇒ 00:43:29.589 Samuel Roberts: Well, I don’t know. With Casey being out, I’m a little more… cautious, because I don’t.
537 00:43:29.590 ⇒ 00:43:30.240 Mustafa Raja: Yeah.
538 00:43:30.240 ⇒ 00:43:30.880 Samuel Roberts: You know.
539 00:43:31.460 ⇒ 00:43:32.939 Mustafa Raja: Yeah, yeah, it makes sense.
540 00:43:32.940 ⇒ 00:43:36.480 Samuel Roberts: Definitely have more of a timeline next week, when we can communicate that to them.
541 00:43:36.670 ⇒ 00:43:37.160 Mustafa Raja: Okay.
542 00:43:38.030 ⇒ 00:43:39.070 Mustafa Raja: Okay, so maybe…
543 00:43:39.070 ⇒ 00:43:42.340 Samuel Roberts: I’m still wondering about moving the database, I’m still wondering about moving…
544 00:43:42.730 ⇒ 00:43:45.040 Mustafa Raja: Oh, we want to move database.
545 00:43:45.620 ⇒ 00:43:58.210 Samuel Roberts: Well, if we want to move it onto their infra, we may be getting on… you know what I mean? We might just use a Postgres database somewhere there. But that’s… this is… I’m still catching up a little bit on the conversations with Tim and stuff.
546 00:43:58.330 ⇒ 00:43:59.110 Samuel Roberts: I don’t know.
547 00:43:59.940 ⇒ 00:44:05.499 Samuel Roberts: But I don’t want to commit to anything until we know this, but I know we’re going to need to start committing at some point soon, so…
548 00:44:06.660 ⇒ 00:44:07.389 Mustafa Raja: Okay, okay.
549 00:44:07.390 ⇒ 00:44:11.850 Samuel Roberts: But I still want to, like, hash a few of these things out before we really say, like, here’s the deadline kind of thing.
550 00:44:13.430 ⇒ 00:44:15.280 Mustafa Raja: Yeah, makes sense.
551 00:44:15.860 ⇒ 00:44:20.960 Samuel Roberts: And especially with Casey out, I mean, even more, because he knows, he has, like I said, more of the knowledge in his head about some of this, but…
552 00:44:22.180 ⇒ 00:44:24.840 Samuel Roberts: Yeah, I would say this is a good path forward for now.
553 00:44:25.390 ⇒ 00:44:28.289 Samuel Roberts: This is gonna have to happen either way, and then…
554 00:44:29.160 ⇒ 00:44:31.989 Samuel Roberts: We’ll see how it looks, and if it’s working better, and if it’s…
555 00:44:32.450 ⇒ 00:44:35.330 Samuel Roberts: Excuse me, as… as efficient.
556 00:44:35.750 ⇒ 00:44:38.409 Samuel Roberts: Because the other side of it is the,
557 00:44:39.070 ⇒ 00:44:42.150 Samuel Roberts: The query lookup tool, the… for the database?
558 00:44:43.740 ⇒ 00:44:48.969 Samuel Roberts: The SQL generation and stuff is kind of slow still.
559 00:44:49.100 ⇒ 00:44:50.720 Samuel Roberts: So there’s other places we can…
560 00:44:51.120 ⇒ 00:44:52.299 Samuel Roberts: I think it’s over, like, 10 seconds.
561 00:44:52.300 ⇒ 00:44:53.360 Mustafa Raja: That’s the reason?
562 00:44:53.500 ⇒ 00:44:56.760 Mustafa Raja: Yeah, part of the reason was, what’s it called?
563 00:44:57.050 ⇒ 00:45:06.299 Mustafa Raja: The… when we queried, so it lives on windmill, right? So Windmill has a… has a cold startup time, no?
564 00:45:06.870 ⇒ 00:45:09.600 Samuel Roberts: That’s what I’m saying, I think we can move some of that into here now, probably.
565 00:45:09.600 ⇒ 00:45:10.649 Mustafa Raja: Yeah, yeah, yeah.
566 00:45:11.070 ⇒ 00:45:17.819 Samuel Roberts: So I think we can start… there’s other places… like, I don’t want to touch that worry… I don’t want to worry too much about it yet, but at some point we can make that move, and then hopefully that’ll…
567 00:45:18.330 ⇒ 00:45:21.019 Samuel Roberts: Make this a little bit faster overall, too.
568 00:45:22.450 ⇒ 00:45:23.029 Mustafa Raja: Yeah, yeah.
569 00:45:23.030 ⇒ 00:45:28.020 Samuel Roberts: But I want to be, like, replacing pieces at one… like, one at a time, rather than, like, just the whole thing at once.
570 00:45:28.180 ⇒ 00:45:29.660 Samuel Roberts: I didn’t want to touch that yet.
571 00:45:31.330 ⇒ 00:45:32.760 Mustafa Raja: Yep, makes sense.
572 00:45:33.300 ⇒ 00:45:34.080 Samuel Roberts: Okay.
573 00:45:34.570 ⇒ 00:45:37.830 Samuel Roberts: Is this, like, the rest of your day here, or is there other things?
574 00:45:39.550 ⇒ 00:45:46.220 Mustafa Raja: I have, some… some default work, too. That just came up, so…
575 00:45:46.470 ⇒ 00:45:46.830 Samuel Roberts: Okay.
576 00:45:46.830 ⇒ 00:45:54.679 Mustafa Raja: But I’ll definitely, by, before, Monday’s stand-up, I’ll have something.
577 00:45:54.680 ⇒ 00:45:55.080 Samuel Roberts: Okay.
578 00:45:55.080 ⇒ 00:45:58.090 Mustafa Raja: I really want to work on this stuff, because I miss it a lot.
579 00:45:58.630 ⇒ 00:46:00.660 Samuel Roberts: Yeah, I know, this is the cool stuff, I get it.
580 00:46:00.840 ⇒ 00:46:02.049 Samuel Roberts: Yeah. I get it.
581 00:46:02.970 ⇒ 00:46:10.849 Mustafa Raja: So, yeah, this is getting me excited. I get to be in my world.
582 00:46:10.850 ⇒ 00:46:12.239 Samuel Roberts: Yeah, I get it.
583 00:46:12.350 ⇒ 00:46:13.579 Samuel Roberts: I’m the only there.
584 00:46:14.050 ⇒ 00:46:25.349 Mustafa Raja: Yeah, so, okay, so this is concluded. I had, I had, one more thing, that I wanted to discuss. So, our…
585 00:46:27.370 ⇒ 00:46:36.390 Mustafa Raja: You do know that we… the decks that we have, are, are being lag… are lagging a lot, right?
586 00:46:38.470 ⇒ 00:46:42.799 Mustafa Raja: By the server that we did deploy on Heroku.
587 00:46:42.800 ⇒ 00:46:43.120 Samuel Roberts: Yeah.
588 00:46:43.120 ⇒ 00:46:44.979 Mustafa Raja: I think it’s just the server.
589 00:46:45.130 ⇒ 00:46:52.919 Mustafa Raja: Part of the… Hmm… because, because, prior to this,
590 00:46:53.570 ⇒ 00:47:05.499 Mustafa Raja: before directly, having, Hiroku serve it, what we were doing is, we had Cloudflare serve it, and that was flawless.
591 00:47:06.610 ⇒ 00:47:07.709 Samuel Roberts: And why do we move?
592 00:47:08.860 ⇒ 00:47:12.659 Mustafa Raja: Cloudflare pages has a limit of 25 MBs.
593 00:47:13.050 ⇒ 00:47:14.170 Samuel Roberts: That’s right, okay.
594 00:47:14.960 ⇒ 00:47:21.449 Mustafa Raja: But now… but now, since Hannah has found a tool that could, you know, shrink
595 00:47:21.450 ⇒ 00:47:35.079 Mustafa Raja: PDFs to, 6 to 7 times. Maybe I could ask her if it’s possible to get all of them below 25, and we could move back to Cloudflare pages?
596 00:47:37.090 ⇒ 00:47:43.289 Samuel Roberts: Maybe. I feel like there’s got to be a better solution here, because we want to get off of GitHub anyway soon, right?
597 00:47:43.540 ⇒ 00:47:46.540 Mustafa Raja: Yeah, yeah, I know we had that discussion also.
598 00:47:46.540 ⇒ 00:47:51.729 Samuel Roberts: Yeah, so I’m trying to think if there’s a better way, like… I mean, S3 might be…
599 00:47:53.630 ⇒ 00:47:57.960 Samuel Roberts: better, and then we can serve them through the platform, even? And then…
600 00:47:58.830 ⇒ 00:48:02.039 Samuel Roberts: I don’t know. I have to think about that a little bit more, but I see what you’re saying.
601 00:48:02.040 ⇒ 00:48:02.920 Mustafa Raja: So, I…
602 00:48:02.920 ⇒ 00:48:06.710 Samuel Roberts: If that… if that compression is a good enough stopgap or not.
603 00:48:07.100 ⇒ 00:48:09.900 Samuel Roberts: But I think there’s still a chance we’re gonna have bigger files, but…
604 00:48:10.610 ⇒ 00:48:17.990 Mustafa Raja: Yeah, I’ll, I’ll send her, send her a message, ask her, if this is something that she thinks is…
605 00:48:18.090 ⇒ 00:48:21.650 Mustafa Raja: Feasible, and if it is, maybe it’s worth giving a try.
606 00:48:22.310 ⇒ 00:48:22.990 Samuel Roberts: Okay.
607 00:48:23.100 ⇒ 00:48:27.619 Samuel Roberts: Yeah, I think it could be, but I’ll definitely give some more thought to the…
608 00:48:27.830 ⇒ 00:48:33.910 Samuel Roberts: moving things into… because I already… I want them to be able to upload from the marketing assets page directly.
609 00:48:33.910 ⇒ 00:48:34.300 Mustafa Raja: Yes.
610 00:48:34.300 ⇒ 00:48:34.799 Samuel Roberts: to be S3.
611 00:48:34.800 ⇒ 00:48:43.849 Mustafa Raja: Okay. Yes, because, the other day, Hannah had, the, the GitHub, GitHub started acting up,
612 00:48:43.850 ⇒ 00:48:44.899 Samuel Roberts: Yeah, it’s so frustrating.
613 00:48:44.900 ⇒ 00:48:46.070 Mustafa Raja: Allow her…
614 00:48:46.240 ⇒ 00:48:53.699 Mustafa Raja: Yeah, allow her to, you know, change branches, and not even discard the changes that were in there.
615 00:48:54.070 ⇒ 00:48:55.369 Samuel Roberts: Oh my god, yeah.
616 00:48:55.370 ⇒ 00:49:07.290 Mustafa Raja: they said was, you know, let’s delete the repository and clone it again, and that will fix. So that fixed. And my instant thought there was, if there were.
617 00:49:07.460 ⇒ 00:49:13.770 Mustafa Raja: A way to directly upload via the platform that’s just…
618 00:49:14.320 ⇒ 00:49:17.980 Mustafa Raja: For marketing, people are also really nice, right?
619 00:49:17.980 ⇒ 00:49:22.730 Samuel Roberts: It’ll be so much better, yeah. I would say… message Hannah, And then…
620 00:49:22.840 ⇒ 00:49:31.360 Samuel Roberts: we can also talk to Gabe about getting that in the plan for internal work soon, because if we can just carve out some time, I don’t think it’ll be very hard to do.
621 00:49:31.630 ⇒ 00:49:33.979 Samuel Roberts: At least the upload. The…
622 00:49:34.160 ⇒ 00:49:40.610 Samuel Roberts: The redirect stuff, we have to think a little more about, but at least we can get off of GitHub to start pretty easily, I think.
623 00:49:41.140 ⇒ 00:49:41.930 Mustafa Raja: Yes.
624 00:49:42.180 ⇒ 00:49:43.210 Mustafa Raja: Okay. Okay.
625 00:49:43.490 ⇒ 00:49:45.199 Mustafa Raja: Okay, this was nice, thank you.
626 00:49:45.670 ⇒ 00:49:48.979 Samuel Roberts: Yeah, no problem. Glad we got to… glad we got to chat.
627 00:49:48.980 ⇒ 00:49:49.530 Mustafa Raja: Yep.
628 00:49:50.420 ⇒ 00:49:55.179 Samuel Roberts: Alright, keep me updated with this, even if it’s over the weekend, shoot me a message or something, but…
629 00:49:55.550 ⇒ 00:49:56.470 Mustafa Raja: Oh, okay.
630 00:49:57.050 ⇒ 00:50:02.980 Samuel Roberts: Yeah, I’m curious to see how it works, because I really want to see this more advanced monster stuff, like, different flows and workflows and things, so…
631 00:50:03.130 ⇒ 00:50:03.859 Samuel Roberts: Let me know if I can help.
632 00:50:03.860 ⇒ 00:50:04.870 Mustafa Raja: Alright, have one.
633 00:50:05.120 ⇒ 00:50:09.609 Mustafa Raja: I wanted to… I wanted to be valuable, too.
634 00:50:09.610 ⇒ 00:50:11.780 Samuel Roberts: Yeah. No, you’re good, you’re good.
635 00:50:11.780 ⇒ 00:50:13.020 Mustafa Raja: Okay, thank you.
636 00:50:13.020 ⇒ 00:50:14.459 Samuel Roberts: Alright, yep, have a good one.
637 00:50:14.950 ⇒ 00:50:16.029 Mustafa Raja: Yeah, bye.
638 00:50:16.410 ⇒ 00:50:16.970 Samuel Roberts: Bye.