Feat: Ragtime! Add RAG capability to stack! (WIP!)

[mbilokonsky]

This is a WIP commit so far, represents current state of this thing but is not yet tested.

Things to note:

1. piggypacking onto the pinecone vector store
2. currently I'm using separate index names for different pinecone use cases, 'concepts' vs 'rag' - we may want a single index with namespaces.
3. this RAG is currently hard-coded to fit the schema of the SQL dump provided, we can tune that.
4. I'm using unusually large chunk sizes, because we are using a large context window. We may not want that.
5. What ultimately gets returned is a queryEngine, which allows you to ask questions and it'll route the question through the whole pipeline. This would be an alternative way to query LLMs.

Note that right now we're hard-coded to OpenAI, but we could trivially replace that with Anthropic etc if we wanted to.

Next steps for me as of current commit: I need to actually run this and make sure it behaves in the ways that I expect. I want specifically to confirm that it persists as anticipated in Pinecone, and that re-running it is idempotent and doesn't rebuild the index on every operation - the docs weren't entirely clear for me in that regard.

But, this should serve as a sort of template for RAG-style work within CB. We can take this in various directions - we could hard-code different rags for different sources, we could parameterize this and make the whole thing filter on metadata so a given query only pulls from specific sources, we could merge this into a single index, etc. There are also a ton of third-party libs that integrate with this, so from this baseline we can really expand in any direction.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Updated (UTC)
course-builder-docs	✅ Ready (Inspect)	Visit Preview	Mar 27, 2024 0:21am
course-builder-poc	✅ Ready (Inspect)	Visit Preview	Mar 27, 2024 0:21am