DOCS Document Intelligence Sandbox

Prototype · Phase 1 · Static

Two manuals.
One patch bay.

A document intelligence workspace — upload PDFs, and it chunks, indexes, and cross-references them automatically. Search across everything at once, find where documents agree or contradict each other, and ask questions that get answered strictly from what's actually in your files, with citations back to the source page.

Library

Every document loaded into this sandbox, chunked on ingest.

Upload

Add a PDF to the library. Text is extracted directly in your browser (nothing is sent anywhere until you click Upload), then chunked and stored.

Works best on text-based PDFs. Scanned/image-only pages won't extract usable text without OCR first — if a document comes back with zero chunks, that's likely why.

Uploaded documents

Everything currently in the database. Deleting here removes the document and its chunks permanently.

Patch Bay

Pick two documents to compare. Terms that surface meaningfully in both get patched together like a switchboard — click a jack to route the connection.

vs

Choose two documents above to find terms they share.

Concept Graph

Every significant term across both manuals, mapped by how often they appear together. Teal nodes are concepts that surface in both documents — the connective tissue of the whole corpus. Click any node to see where it actually appears.

Appears in both Document-specific Document-specific · Cross-document link Same-document link

Click a concept above to see every passage that mentions it, across both manuals.

Process Flow

Turn a topic into an ordered, cited sequence of steps reconstructed from both manuals — the same mechanic that, on a case file, would rebuild a chain of custody or an approval path from scattered filings.

Drafting aid, not legal or technical advice — verify every step against the source before relying on it.

Brief Generator

Turn a topic into a structured briefing memo — overview, key definitions, cross-references between the documents, and open questions a reviewer should still chase down.

Drafting aid, not legal advice — every citation should be checked against the original document before you rely on it.

Chunk Explorer

How each document was split for retrieval. Every segment below is a real chunk with page provenance — hover to inspect.

Hover or tap a segment to preview that chunk.

AI Insights

Retrieval meets generation, live. Every answer below is genuinely generated by Claude from real retrieved passages — nothing here is pre-written.

Drafting aid, not legal advice — treat everything below as a starting point for review, not a conclusion.

Ask across documents

Live

Ask a question that spans your uploaded documents. This retrieves the top matching chunks and asks Claude to answer strictly from them, with citations back to page numbers.

Story Mode

Live

Turns a cross-referenced cluster of chunks into a plain-language narrative — the same mechanic that, on a legal document set, would turn a folder of filings into a coherent case timeline.

Conflict Finder

Live

Instead of answering a question, this actively hunts for places where your documents define, describe, or handle the same thing differently — contradictions, terminology drift, or gaps between them. On a real case file, this is how you'd find where a witness statement conflicts with a contract, or where two filings disagree.

Devil's Advocate

Live

Give it a claim or a position, and it looks for the strongest grounded pushback against it — genuine counterpoints from the documents, not manufactured ones. On a real case, this is how you stress-test your own argument before the other side does.

How this works

Each request retrieves the best-matching passages from both manuals (same mechanism as the Search tab), sends only those passages to Claude with instructions to answer strictly from them, and returns a real generated answer with page citations. Nothing is pre-written below this line.

Binder

Collect excerpts from anywhere in this site — Search, Chunk Explorer, Patch Bay, Concept Graph, or any AI Insights citation — into one ordered bundle, then export it as a real PDF. On a real case file, this is how you'd assemble an exhibit bundle from scattered documents.

The exported PDF is a compilation of excerpts you selected, with citations — it is not a legal filing and hasn't been reviewed for accuracy or completeness. Verify everything against the original documents.

Nothing added yet. Look for "＋ Add to Binder" wherever you expand a passage — Search results, Chunk Explorer, Patch Bay, Concept Graph, or any citation chip.