HYBRIDGE

Enterprise Search

Ask the company anything. Get answers with receipts.
An internal assistant that answers from Hybridge documents, cites every claim, and respects who is allowed to see what.

What it is in one breath
Where it runs and who can use it

Live demo · hybridge-search.web.app

The product

Two tabs, one loop

Ask Hybridge: a chat that streams answers grounded in company documents. Click a citation and the source opens at the exact highlighted passage.
Dropoff: authorized people feed the knowledge base by dropping documents into a vortex, scoping who can find them, and watching ingestion status.
The loop: knowledge in through one governed door, answers out with sources, always.

The write path

How knowledge gets in

→ → → → → →

The original file is kept safe and always downloadable, permission checked.
Failures are honest: unparseable documents land in a Failed state with the reason shown.

Metadata

Every document gets a small card

Gemini proposes a title, summary, entities, and tags. The uploader confirms the governance fields. The card stays small no matter how big the document is.

Who can find it

Everyone, or specific groups. Empty means open to all of Hybridge.

How much to trust it

Document class drives an authority score, computed on the server.

When it applies

An effective date powers recency tiebreaks when sources disagree.

The read path

Finding the answer

→ ＋ → → → →

Meaning search catches paraphrase: "time off rules" finds the PTO document.
Keyword search catches exactness: part numbers, "35 Ncm", names.

Access control

Not everyone sees everything

Users belong to groups; documents are scoped to groups at drop time.
Deny by default, enforced inside the database query itself.
Access comes only from your verified identity, never from the request.

Trust

When sources disagree, we never flip a coin

Superseded versions are dropped first. Retrieval only sees current documents.
Higher authority wins; ties break to the newer effective date.
Real disagreements are said out loud, with both sources cited.

Trust

Grounded, or silent

"Answer only from the retrieved documents. Cite every claim.
If it is not in the knowledge base, say so."

No retrieval means the model is never called. No general knowledge answers, ever.
Every claim carries a citation that opens the exact passage of the exact version.

Growth

Built to grow, not preloaded

A new document is searchable minutes after drop. Nothing gets rebuilt.
Re uploads are safe: duplicates skipped, revisions become new versions.
Honest empty states tell us exactly what to ingest next.

Under the hood

How it is built

All Google Cloud

Cloud Run, Postgres with pgvector, Cloud Storage, Identity Platform, Vertex AI. No third party SaaS.

One database

Vectors, keywords, permissions, versions, and chat history in one Postgres.

Engineering discipline

Terraform everything, versioned deploys, 122 automated tests, adversarial review.

Considered and rejected

Vectorless LLM navigation and graph extraction. Hybrid won on latency, cost, and fit.

Roadmap

What is next

Live answers switched on in production.
Near term: chat attachments, web search toggle, Google Drive sync, multi turn memory, batch uploads, a document library.
Later: voice, dark mode, usage analytics, continuous evaluation, scale hardening.

Live demo

Watch it work

Drop a policy into the vortex and scope it to a group.
Ask the question it answers; watch the answer stream with citations.
Click citation 1: the source opens on the exact highlighted passage.
See two documents disagree, and the answer explain which one it trusted.
Ask from outside the group: the document simply does not exist for them.

hybridge-search.web.app · questions and feedback welcome