AgeBridge Guided ProjectEvidence attached

Internal Knowledge Assistant

A private assistant that answers internal questions with links to the source, and never shows a document the asking person could not open themselves.

NB

Noa Barak

Applied AI engineer

Business problem

A 140-person company kept policies in Drive, specs in Confluence and decisions in Slack. New joiners asked the same questions for months, and half the answers they got were out of date.

Solution

A private assistant that answers internal questions with links to the source, and never shows a document the asking person could not open themselves.

What was built

  • •Nightly sync from Google Drive, Confluence and selected Slack channels
  • •Access rules stored alongside every indexed chunk
  • •Retrieval filtered by the asking user's permissions before ranking
  • •Answers that always link back to the source document
  • •Thumbs-down feedback with a weekly review of failures

Architecture / workflow

  1. 1Connectors sync documents and their permissions nightly
  2. 2Each chunk is stored with the group IDs allowed to read it
  3. 3A query is filtered by the asker's groups before ranking
  4. 4Answer returned with source links the asker can actually open
  5. 5Negative feedback is queued for a weekly review

Integrations

Google DriveSlackConfluence

Tools / stack

OpenAIPythonPinecone

Screenshots

Screenshot slot
Screenshot slot

Demo

Same question asked by two people with different access

Results

No verified results shared. AgeBridge only displays real, builder-reported outcomes.

Evidence

  • Architecture notes · Permission-aware retrieval designHow data flows and why you designed it that way.
  • Screenshots · Restricted query from two accountsReal screens from the build, not mockups.

Proof

Proof score

25/100

Add demo video — 60 seconds showing the thing actually running. The single strongest proof.

Not reviewed yet. Evidence is shown exactly as the builder attached it.

Decisions and tradeoffs

  • •Permissions are enforced at retrieval, not by asking the model to be discreet. A model instruction is not an access control.
  • •Stale content is a bigger risk than missing content, so every answer shows the source's last-updated date.
  • •Slack is indexed selectively. Most channels are noise and poison the index.

What went wrong

The first sync flattened permissions and briefly made an HR folder searchable internally. It was caught in testing with two accounts, which is now step one of the build.

What I'd do next

Add an evaluation set of 50 real questions so changes can be measured, not guessed.

Need something like this?

Talk to the builder behind the project, or build your own version.