Loading
Retrieval-augmented generation · intermediate
Look up the right facts first, then let the model answer from them.
The main tools by job. Tap one to see what it does, then learn it here.
Models
Frameworks
Vector stores
Parsing documents
Embeddings
Reranking
Evaluation
Running open models
The scenario
Your support bot retrieves the top 5 chunks. On hard questions the passage with the answer sits around rank 15. Raising the count to 30 makes answers slower, pricier and, oddly, worse.
The question
How do you get the right passage to the model without stuffing the prompt?