AI Build Week – Day 3: How we made RetoolGPT

📣 Retool Announcements 🤗 Community Happenings
1

Welcome to Day 3 of ai-build-week!
We’re halfway through :dart: and loving all the sharp questions and creative energy so far. Keep it coming!

If you missed a session, catch up here:
:point_right: Day 1: Prompt Arena Showdown
:point_right: Day 2: Vibes-Based Search with Vectors

Today we’re going behind the scenes of RetoolGPT with @keanan — and unpacking how we built our embedded AI assistant for docs and support using Workflows, Agents, and smart prompt design.

:alarm_clock: Live at: 2025-07-09T17:00:00Z→2025-07-09T18:00:00Z — right here in the forum
:speech_balloon: Say hey if you’re tuning in! We’ll be hanging out and answering questions throughout the day.

What you’ll learn:

  • How RetoolGPT is built with Workflows + Agents
  • How to structure prompt chains and select LLM's
  • Behind-the-scenes lessons from building AI into support and documentation flows

Get involved:
:file_folder:Download today’s resources
:speech_balloon: What’s one takeaway or “aha” moment you’ve had so far this week? Share it in the thread!
:tada: Midweek swag drop!
Congrats to @Rik_de_Rooij and @Shawn_Optipath — winners of Day 1 & 2 of AI Build Week! :billed_cap:

Keep the great questions coming… more prizes on deck :eyes:

:wave: Welcome back @ai-build-week-25 — drop a hello, share a takeaway, or ask your questions below!

← Back to the full AI Build Week schedule

6 Likes
2
How many AI Build Week sessions are you planning to join?
  • Just 1 – popping in for a taste :eyes:
  • 2 – double feature :clapper::clapper:
  • 3 – I’m in my AI era :robot:
  • 4 – basically co-hosting at this point :microphone:
  • All 5 – full week, let’s gooo! :rocket:
0 voters
2 Likes
3

I just caught up on the recording. Really great session! :clap:

I had a question about something mentioned in the demo. It relates to using branches in workflows and the retool trigger data (such as searchConfluence == true) to decide which Vectors to search:

CleanShot 2025-07-10 at 12.59.27@2x
CleanShot 2025-07-10 at 12.59.27@2x1920×1151 79.1 KB

This feels like it wouldn't scale well. If you've got X tools, you need to create X branches, which trigger X Vector Searches, which then need to be merged for the AI prompt.

I was wondering if you could share more about the advantages of this approach vs storing all of the documents from all tools in one Vector, and using filters for the Vector to decide which "sources" to include in the query? Unless I'm mistaken about the Vector Search functionality, using the filters would mean we don't need branches and we could rely on a single Vector query.

I hope the question makes sense, I'm new to working with Vector Search and LLMs. :laughing:

1 Like
4

the Select Vectors field can be a function if you click the little fx to the right of the dropdown button. So you could write some simple JS to select the vectors based off the tool. this would let you take your tool list and use it as input for a loop that contians query1. the loop now does a vector search for each tool using the proper vector stores.

1 Like
5

i forgot to answer your question, sorry. you could do it that way using filters, but you normally want everything in the vector to be relevant in some way. I think the biggest difference is having to remember to update your filter list which is on a different screen than the workflow vs updating the logic to include a new vector store.

you could run into a problem where some files with filters you don't include are actually related but don't get included because the majority of files under that filter are unrelated so you either don't know the file is related and don't include it or don't include it because 99% of the filter doesn't need to be included.... either way you could be filtering out an important file on accident due to forgetting to update filters. forgetting to add a whole vector store is a bit more difficult to do and easier to debug :innocent:

1 Like
6

Thanks @bobthebear! I appreciate you taking the time to write such a detailed explanation. Totally makes sense that multiple vectors would be the safer option to ensure no relevant files are excluded. I can see one large vector store becoming very difficult to manage as it scales. I'll likely default to one Vector store per source if/when I take on this kind of project.

Thanks again! :raised_hands:

2 Likes
7

:bulb: AI Build Week: Day 3 Q&A Recap

Thanks to everyone who joined live and submitted questions during the session!

:chart_with_upwards_trend: Fun fact: Day 3 brought our biggest burst of technical questions yet — from Slack API changes to multi-key Agent setups. Thanks to everyone who joined live and shared real-world scenarios!

Here’s what we covered:

:sparkles: Building RetoolGPT

Q: Is the RetoolGPT workflow available to review?
A: Yes! You can download the full JSON export and import it into your own instance. Here’s the resource folder.

Q: What does RetoolGPT use under the hood?
A: It's powered by Workflows, Agents, Retool Vectors, and smart prompt chaining. We also use a modular component pattern (MCP) to keep things maintainable. Failover logic ensures the system "fails loud" so we can improve it over time.

Q: Can we build something similar for internal copilots?
A: Absolutely. You can build intelligent copilots by calling Workflows, embedding apps, or triggering Agents inside your internal tools. Retool gives you a lot of flexibility to tailor the experience to your use case.

:open_file_folder: Vectors, Storage & Chunking

Q: Do Retool Vectors share the same 5GB limit as Retool DB?
A: Yes. Vector storage is part of your Retool DB instance. If you need more than 5GB, you'd need to expand your DB storage. We haven't seen many customers hit this cap yet.

Q: Can I use other vector databases like Pinecone?
A: Not natively right now. You can integrate external vector DBs like Pinecone using REST API resources. Submit a feature request if you'd like to see more native vector DB support.

Q: How does Retool handle chunking for vector ingestion?
A: Retool performs chunking internally, but you can pre-process your data using tools like RecursiveCharacterTextSplitter to customize chunk size and overlap before ingestion. If you’d like more control over chunking settings, let us know.

Q: Can we ingest Confluence docs into Retool Vectors?
A: Yes, as long as you can extract your Confluence content, you can convert it to a format Retool can ingest and use it in your workflows.

Q: How do vectors affect token usage?
A: Adding vectors means your model receives more context, which increases token count and cost. It's a tradeoff for better relevance and response quality.

Q: Can we schedule vector ingestion and prevent overlap?
A: Yes. You can schedule ingestion with Workflows and add logic to skip runs if a prior one is still active. This helps avoid concurrency issues.

:rocket: Agents, API Keys & Extensibility

Q: Do we need to bring our own API keys for Anthropic, OpenAI, etc.?
A: Retool provides managed credentials for OpenAI and Anthropic with reasonable usage limits. You can also bring your own keys. For other models like Deepseek, you’ll need to supply your own credentials.

Q: Can different apps use different API keys?
A: This isn’t supported today, but our team is actively working on better multi-key management and permissioning at the platform level. Submit a feature request if you'd like to prioritize this.

Q: Can Retool handle multi-step agents without Make.com or external tools?
A: Yes! You can chain Workflows, use Agent tools, and pass context all within Retool. If you want to integrate with Make.com, that's also possible via API calls.

Q: How do Slack’s new LLM API terms affect integrations?
A: Slack's changes mostly relate to historical data ingestion. Triggering workflows or bots in real time should still work fine, but it's good to double-check their latest terms.

Q: Can we use Gemini’s "Grounding with Search or Maps" feature in Retool?
A: This isn’t supported via the native AI action. To use Gemini’s grounding capabilities, you'd need to create a custom API integration. Submit a feature request if you'd like native support.

:toolbox: Miscellaneous

Q: What’s the case for building this vs. using tools like Guru or NotionAI?
A: Tools like Guru and NotionAI are great for general use, but building with Retool gives you full control over logic, model behavior, and integration with your internal systems.

Q: Are there lists of available actions for each AI component?
A: The best way to discover them today is through the UI and docs. Actions like "generate text" and "generate chat response" are most common. Copilot-style helpers will improve this experience in the future.

Q: Can we use Office docs stored on Egnyte instead of Google Docs?
A: Yes — if Egnyte offers an API (it does), you can connect it to Retool and pull in your document data.

💬 What did you think of Day 3?

Drop a comment — or just vote in the quick poll below:

  • :green_circle: Loved it — more like this please!
  • :yellow_circle: It was solid, but I had questions.
  • :red_circle: Not quite what I expected.
0 voters
2 Likes
8

Hello guys!
Well done with the Retool GPT, I LOVED IT!

I've been tryin to replicate the ChatGPT interface for sometime now!
I noticed that within the Resources folder only contains the Workflow JSON, will the App JSON be shared or that's not on the plan.

Thank you! and amazing work!

9

Hey @Infinitybht - thank you so much... we’re so glad you enjoyed it! :raised_hands:

At the moment, we’re only sharing the Workflow JSON as a behind-the-scenes peek. The full RetoolGPT app isn’t planned for release, since it includes some internal patterns and UI components that aren’t quite ready to be bundled into a reusable template.

That said, we’re definitely happy to help you build something similar! The workflow is a great starting point, and feel free to drop your questions or progress in the Community. We’d love to see what you create. :star_struck:

2 Likes
10

Thanks Sarah! (beautiful name btw!)

I promise to build a great GPT App and share it with the community here!

Wait for my response guys!!

2 Likes