Grounded · Verified · Traced

Answers you can actually cite.

A self-correcting retrieval pipeline reads your documents, rewrites its own query when the evidence falls short, and has a second, independent model verify every answer before you see it.

Ask a question about your documents…
The pipeline
  1. 01

    Retrieval

    Dense semantic vector search over your own corpus.

  2. 02

    Self-correction

    Rewrites its own query until the retrieved context is sufficient.

  3. 03

    Verification

    A second, independent model checks every answer for faithfulness.

  4. 04

    Citation

    Each claim traces back to the specific passage it came from.

Trace of a single query

“How does scaled dot-product attention work?” · 2.4s · 6 sources

  1. 01Retrieved candidate passagespass 1 · 0.4s
  2. 02Rewrote query — context insufficientself-corrected
  3. 03Retrieved candidate passagespass 2 · 0.9s
  4. 04Critic checked answerabilityself-check · passed
  5. 05Synthesized grounded answer6 sources · 1.1s

Claude Haiku 4.5 answered · GPT-5 nano verified independently

Agreed · 0.94

Ask it something your documents can answer.