[Feature] RAG for Extraction (SQLite).

[enoch3712]

Sometimes in large files, extraction will contain a lot of extra information that will be in 1% of the pages. To solve this problem, the user can add RAG to solve this problems and keep the token usage low

[zeweihan]

Great feature request! We've been tackling a similar challenge in AI WorkDeck — our approach has been to integrate document-level RAG at the workspace level, where users can index and retrieve relevant sections before running extraction contracts. This keeps token usage low while maintaining extraction quality across large documents.

A few things we've found useful:

• Chunking documents semantically (by sections/headings) rather than by fixed page boundaries improves retrieval relevance significantly
• Pre-filtering with lightweight embedding models before sending to the LLM extractor reduces costs by ~60% on large legal/technical documents
• Storing extraction metadata alongside chunks helps with incremental re-extraction when documents are updated

Would love to collaborate or share more details on our approach if it's helpful for ExtractThinker's RAG integration. The ORM-style extraction model you've built is really elegant — it'd be exciting to see it work with a retrieval layer.