Skip to diagram
Clarity.Explore the demo

HOW CLARITY WORKS

A clearer story. A better first meeting.

Clarity helps clients prepare the facts so their lawyer can focus on legal judgment.

  1. 1 Client

    Tells the story

    Explains what happened and uploads supporting documents.

  2. 2 Clarity

    Asks & organizes

    Asks factual questions and links information to its evidence.

  3. 3 Client

    Checks & confirms

    Reviews the draft and confirms what will be shared with the firm.

  4. 4 Lawyer

    Reviews & decides

    Receives a sourced brief, checks the evidence and applies legal judgment.

The lawyer can ask the client for clarification.

The draft stays hidden from the firm until confirmation.
Then Clarity prepares a brief linked to the confirmed evidence for human review.

Your words. Linked evidence. Human judgment.

2.54M

indexed legal passagesA dated collection, not all French law.

32s

to prepare a brief in one synthetic testOne measured run, not a speed guarantee.

Technical details & measurement sources

Connected backend capabilities. The existing live preparation service uses Mistral OCR for PDFs and images; text and DOCX text are parsed directly. After confirmation, it can send the fixed submission and permitted legal passages to Mistral Large 2512. Quotes must match stored source text. The resulting brief is internal, marked unreviewed and requires human review. These capabilities depend on the configured backend and are separate from this pitch’s prepared final result.

Search in that backend. BGE-M3 embeddings on the DGX feed PostgreSQL/pgvector. The preparation route combines exact vector search with keyword search, with optional Jev reranking. It uses exact search, not the available HNSW index. Up to 15 passages are retained and up to five parent documents are expanded within a size budget; up to 10 eligible public legal passages join the confirmed client evidence for generation. Citation-graph expansion is not implemented. Integration code

Measurements, 4 October 2026. The corpus audit records 2,542,770 unique passages across nine public-source snapshots; standalone Judilibre and JORF are not included. The timed run took 32.39 seconds on codex/testing, using an isolated local application, remote providers and fictional evidence. These observations do not measure legal accuracy or production performance. Corpus audit · Test receipt

This pitch. Start with Lucy’s pre-filled draft or enter fictional information yourself. In the browser-session demo, final preparation and team assignments use deterministic rules after confirmation. The guided factual questions also follow rules; they are not generated on the spot. The hybrid chat serves a prepared, sourced SAS explanation and can send other general questions through the existing RAG and Mistral services when their connections are configured. An unavailable connection stays visible as an error; it is never presented as a live answer.

Demo boundaries. The original connected application uses a Vercel gateway, a Google Cloud VM with SQLite, the DGX and the Mistral API. The browser-session pitch can run without those connections. All case data is fictional; appointment booking is simulated. This demo does not provide production SSO or real-client onboarding.