Deep File Retrieval
Deep File Retrieval is the retrieval mode for large knowledge bases. It combines hybrid semantic and keyword search with hierarchical retrieval, so your assistant finds the right passage even across hundreds of documents.
Availability: Deep File Retrieval is included with the Enterprise Plan, which supports up to 100 files (up to 10 MB each). Other plans use standard retrieval.
Why standard search falls short
Keyword search alone misses passages that answer a question with different wording. Pure semantic search has the opposite problem — it can gloss over exact identifiers like SKUs, clause numbers, or API names. And with many large files, both approaches start returning plausible-but-wrong passages from the wrong document. Deep File Retrieval addresses all three problems by combining search strategies and narrowing by document structure first.
How it works
Indexing
When you upload a file, it is parsed, split into chunks, and stored in the Qdrant vector database with both dense (semantic) and sparse (keyword) representations.
Hybrid search
Each question runs a dense and a sparse search in parallel. Dense search finds passages that mean the same thing even with different wording; sparse search catches exact terms like product codes, names, and error messages. The results are merged and re-ranked.
Hierarchical retrieval
For large collections, retrieval works top-down: it first narrows down which documents and sections are relevant, then pulls the specific passages. This keeps answers accurate even across hundreds of pages.
Answer generation
The retrieved passages are handed to your assistant's model, which answers using that material — so responses stay grounded in your documents.
Supported content
Deep File Retrieval works with everything the knowledge base accepts:
- PDF, Word, Office, and CSV files
- Scanned documents and images, extracted with OCR
- Website content imported through the built-in web crawler
See How files are organised for upload limits and organization tips.
When it helps most
Technical documentation
Manuals, API references, and specifications where exact terms (part numbers, endpoints) matter as much as meaning.
Legal and compliance
Policies and contracts where an answer must come from the right clause in the right document.
Research collections
Reports and papers where answers need to combine findings from several sources.
Internal knowledge bases
Procedures and training material spread across many files and formats.
Deep File Retrieval is part of the Enterprise Plan. Compare plan limits and features to see if it fits your knowledge base.
Compare plans