Monograph

Nocode Lisboa #10: AI + Nocode

2024-06-21 11:11

Demo the App

High-level Blueprint

Important Concepts

  • The systems involved: Bubble, LangChain / n8n, Vector Store db, LLM

  • What is a Vector Store db and why do we use it?

  • RAG app (Retrieval Augmented Generation) - overcomes limitation of LLM context window

  • Embeddings : Text :: Dewey Decimal System : Books

  • Infrastructure running everything self-hosted: Docker, Google Cloud, Portainer mgmt

  • Value of doing this ^^

  • Value of using middleware like n8n + LangChain

  • Working with REST APIs

App Architecture

  • Front end management UI and display UI in Bubble

  • Process of indexing new content

  • N8n workflow that generates the embedding & creates the SupaBase entry

  • Process for requesting a new batch of intros

  • N8n AI agent that invokes tool to query SupaBase and return results

  • look at rationale of the v2 algo work in progress

Hard-earned Lessons

  • Using modular n8n workflows via webhooks & http posts (ie. why "when called by another workflow" triggers don't work for this)

  • Establishing local vars to the current workflow with "Set"

  • Code block JSON copy/paste as a chokepoint + disable previous workflows for reducing # calls

  • Building up a dynamic prompt using the "Set" node

  • Debugging in n8n with test mode. Executions tab for sleuthing issues historically

  • Navigating using test/live with both n8n and Bubble

  • Postman (or Bruno OSS equivalent) + Pipedream for debugging at the API request level

  • Copy/Paste entire workflows in forums (useful for backups too)

  • Using ChatGPT to write functions for data transformation in Code Nodes

  • Structured JSON output parser and auto-fixing process

Learning Resources

Questions from the event:

  1. What was the real value people got from this app? Couldn't they just meet at the conf?

  2. Isn't this overkill to use autonomous agents for this when you could just do a simple matching algorithm?

  3. What costs are we talking roughly for running all these services each month?

  4. How tough was it to instantiate the proper database in SupaBase and make that work?

  5. How do you decide which pieces to handle in Bubble vs. N8n?

  6. In the prompt engineering piece, what goes in the human message vs. the system message portion?

Projects mentioned to check out:

https://github.com/stanfordnlp/dspy

https://neo4j.com/

https://www.plasmic.app/

Video recording:

https://youtu.be/6b1o_7A2MP0