Give your questions somewhere to look.
Start with fictional practice documents, or add your own. Nothing is sent to AI until you explicitly approve a connected action.
Knowledge library
Up to 50 documents / 1 million characters per project. Replacing or removing a document invalidates its index and results.
Small pieces. A clearer view.
Choose a stage to learn what it does.
Build an index to inspect your passages.
Chunk explorer
Open a chunk to highlight its exact range in the cleaned source. Overlap may be reduced at short boundaries to guarantee progress.
Show me the evidence.
Scores rank passages within this search. They are not probabilities, factuality scores, or proof that a passage answers your question. Reranking is a deterministic 20% term-coverage boost.
Retrieved passages
Your evidence trail starts with a question.
Grounded answer workspace
Extractive mode · no AI generation
Exact context sent to generation
No context yet.
Human review notebook
Change one thing. See what changes.
Compare local keyword pipelines using the current question and metadata filters. Provider tokens/cost are zero; token counts below estimate context length only. Answer quality requires human review.
Good evidence is testable.
Build a verified reference set. Deterministic checks measure source-document recall, reference word coverage in retrieved text, and no-hit abstention. These do not establish answer correctness or groundedness.
Knowledge gap report
Become an evidence detective.
Each mission asks you to change the system, inspect the effect, and explain the result. Progress stays in this browser.
Privacy & optional connection
Documents and projects are stored in this browser using IndexedDB. Anyone using this browser profile can access them. Project separation is logical, not a multi-user security boundary. Back up important work; clearing browser data removes it.
Connected mode requires a server administrator to configure the included PHP Gemini gateway. Each external action asks for confirmation. No credentials are stored in browser storage. No analytics are enabled here.
Not connected. Embeddings and generated AI answers are unavailable.
Retrieved content is untrusted. The generator receives it as evidence, with no tools or ability to change permissions. This reduces impact but does not guarantee immunity to prompt injection. Verify every claim; a valid citation can still be misinterpreted.
Do not upload confidential records or personal information to a third-party model without authorization.