Limitations

We would rather be clear about the current state than oversell it.

Graph answering is off by default

The knowledge graph is built during ingestion and can be used to expand retrieval, but graph-based answering is disabled by default. On our evaluations it did not yet beat plain hybrid retrieval, because automatic relationship extraction — the part that makes the graph dense and correct — is still being improved. We enable graph answering per-tenant only when it measurably helps. Until then you get robust hybrid retrieval with citations.

Retrieval is tuned for English

Extraction, chunking and re-ranking are tuned for English content. Other languages may work but are not yet validated.

No OCR

Scanned or image-only PDFs are refused. Run them through OCR before uploading.

Retrieval quality is corpus-dependent

Answer quality depends on your documents. If a question is answered poorly, the most useful thing you can send us is the exact query, the k used, the answer you got, what you expected, and the relevant document_id — real-document feedback is how retrieval improves.

Availability

The reference deployment is single-node. Backups and tested restores ship with it; high-availability replication is not enabled out of the box — add it before relying on the service for production traffic.

What this means for you

• Your integration is stable: the API is versioned and backward-compatible.
• Answers today are grounded in your documents and cited.
• Graph features are a measured upgrade path, not an unproven default — when they’re turned on for you, it’ll be because they were shown to help.