Ask 200 documents a question. Get the page it came from.
An enterprise knowledge assistant where every streamed answer cites its source — and clicking the citation lands on the exact highlighted passage in the original PDF.
Live demos are password-gated — request access via the contact page and the password comes back in minutes.
Mid-size companies sit on hundreds of policies, manuals, and specs nobody can search — so people interrupt each other for answers that are already written down.
A workspace assistant that streams grounded answers from ~200 indexed documents, with a source viewer that renders the cited page and highlights the exact passage — plus the pipeline dashboard that proves the corpus is real.
- Token-by-token streaming answers with numbered citations
- Citation click → the original page with the passage highlighted
- A document base whose counts reconcile: 200 docs, 191 indexed, 11,857 chunks
The challenge
Enterprise buyers don’t fear wrong answers as much as unverifiable ones. An internal assistant that can’t show its page is a liability in HR and safety contexts.
Knowledge-ops teams also need the unglamorous half: what’s indexed, what failed, what a re-index costs — visible, not buried in logs.
The solution — three decisions
The page is the proof
The source viewer renders a document page — serif, justified, footered — with the cited passage in marker yellow. Answer → page → highlight is the whole trust story.
Streaming is non-negotiable
Answers render token by token. Instant text reads as canned; a stream reads as a live model at work.
Ops get a real surface
Library, ingestion jobs, and retrieval settings (top-K, thresholds, reranker, llama-3.3-70b) are first-class screens — because the buyer here is the team that runs it.
How it works
The demo implements this shape end to end with a simulated service layer — the “extend for production” section lists what swaps in for live deployment.
Product tour
The exact sentence the answer citedDocument, page count, chunk countPage 3 of 18 — browse the source
What the demo shows
- Three scripted question flows across HR, maintenance, and PTO topics
- Ten past conversations that reload their grounded answers instantly
- A workspace settings page with working retrieval controls
- Dense enterprise UI: 13px body, 36px rows, icon-rail shell
Under the hood
- Cited passages are stored per document+page; the viewer renders them into generated page context so every citation resolves
- Non-cited pages get deterministic filler prose per category — browsing never breaks the illusion
- Chunk totals are computed from the doc list; library, KPIs, and jobs can’t drift apart
- Icon-rail shell (the portfolio’s sixth distinct app shell) keeps 200-doc density navigable
Meridian Corp is fictional and labeled as a demo on every screen. Nothing here is presented as client work: no client names, no outcome metrics, no testimonials. The proof is the running product — open the live demo above and check every claim.
What we’d extend for production
- Real embedding + retrieval over the corpus (e5/bge + pgvector)
- Chunk-to-page-coordinate mapping for true PDF highlighting
- Per-space permissions so answers respect document ACLs
Build something like this.
Fixed scope in writing before any money moves, demos during the build, and full code ownership at handover.