RAG

Retrieval-Augmented Generation: chunking documents, embedding them, retrieving relevant chunks, and grounding an LLM's answer in them instead of relying on model memory.

Why we need this / what value this brings

Lets an LLM answer accurately about your own private/current data instead of hallucinating from stale training data.

When to use this

When answers must be grounded in specific documents (docs, listings, policies) that change over time or were never in the model's training data.

How to use or implement this

Chunk documents → embed chunks → store in a vector DB → on a query, retrieve top-k similar chunks → inject them into the prompt as context before calling the LLM.

Research questions

• Chunking strategy: fixed-size vs semantic vs recursive splitting — trade-offs?
• How does retrieval quality get measured (recall@k, MRR) before it's wired into a chat UI?
• When is RAG the wrong tool (e.g. small enough context to just paste the doc in)?

---

Empty folder — drop notes, links, and findings here as you research.