The knowledge base is per tenant. When you upload a document, FAQ, PDF, or scraped URL, FixControl splits it into chunks, embeds each one, and stores them for retrieval. Hybrid retrieval (vector + keyword) at query time gives the agents grounded context per ticket.
User-facing knowledge management lives in the user knowledge guide. This page is the operational picture.
Fingerprinting
Every chunk is stored alongside the provider, model, and dimensions it was embedded with — its fingerprint. If the embedding provider or model changes, a new query embedding can't be compared against chunks embedded under the old fingerprint — they live in different vector spaces.
Retrieval handles this automatically:
- Fingerprint matches → vector search.
- Fingerprint mismatch → falls back to keyword search until those chunks are re-embedded.
The system never breaks on a provider change; it degrades gracefully to keyword search.
Re-embed flow
If your deployment's embedding provider or model changes, you re-embed from the UI:
- Go to Settings → Knowledge → Re-embed all (admin only).
- FixControl re-embeds chunks in the background. While that's running, queries fall back to keyword for any chunk not yet re-embedded.
- The progress bar on the Knowledge page shows re-embedded / total.
If a re-embed pauses partway (provider down, rate limit), it resumes from the last successful chunk. No double-embedding.
Health surface
The admin embeddings health page shows:
- Active provider + model for the deployment.
- Total chunks and how many match the active fingerprint.
- Stale chunks — embedded under a previous fingerprint, search falls back to keyword.
- Last re-embed run with duration, success count, and error count.
If you see stale chunks weeks after a provider switch, run the re-embed again.
Hybrid retrieval
At query time the Support Agent (and other agents that use knowledge) build a query embedding, run a vector search for the top-N candidates, rerank those by keyword overlap, and return the top-K to the agent.
N defaults to 30 and K defaults to 5. Both are tuneable per tenant via the knowledge settings if you want to widen or narrow the funnel.
Tenant isolation
Every knowledge search is scoped to your tenant. A user in another tenant can never retrieve one of your chunks, even via the API. Cross-tenant retrieval is not possible.
FAQ
Does the deployment switching embedding provider require a re-embed? Yes. Different model spaces, different dimensions, different fingerprints. Run the re-embed job after the switch.
Can different tenants use different models? Not today — the provider/model is set per deployment. Let us know if per-tenant providers matter for you.
How much does it cost? For cloud providers, pricing scales with token count; for a 50-document knowledge base the initial embed is typically a few cents. Re-embedding the same content under a new model costs the same again. Self-hosted embedding is free other than your hardware.
Can I exclude some sources from agent context? Yes — set the source's status to reference instead of indexed. It's then visible by name but not searched.