Pradra StudioDocs
Sign in

Knowledge sources

A language model only knows what it was trained on. A knowledge source lets a generative capability answer from your documents — policies, product manuals, past cases — by retrieving the most relevant passages and giving them to the model as context. The pattern is known as retrieval-augmented generation (RAG).

Versioned, like everything else

Knowledge sources live in your workspace and are versioned: re-ingesting updated documents writes a fresh version, and a capability pins the exact version it retrieves from. Updating the source never silently changes what a deployed capability answers — you re-pin deliberately, and can point back just as easily.

The PII safeguard

During ingest, a deterministic scan looks for personally identifiable information (emails, phone numbers, card-like numbers). A source with findings is marked flagged and cannot be retrieved from until a person reviews and acknowledges it. Only the counts are recorded — the matched text itself is never stored in the findings.

Retrieved context is data, not instructions

Passages retrieved from your documents are handed to the model explicitly framed as reference material — they fill a context slot in the prompt, they do not become instructions. A malicious sentence inside an ingested PDF cannot steer the model the way the prompt does. Retrieved text also passes through the content filter like any other input.

Every answer records which chunks it used, so "why did it say that?" has an inspectable answer in the prediction history.

Working with knowledge sources in the Studio

  1. Go to Knowledge sources in the sidebar. Create one: name it, pick the embedding model, set the chunking parameters (the defaults are sensible), and choose whether to scan for PII.
  2. Ingest content. Ingestion runs in the background; the row shows ingesting, then ready — or flagged if the PII scan found something, with an Acknowledge action after review.
  3. Open the generative capability in AI Studio → Retrieval tab: bind the source (at its pinned version), set how many chunks to retrieve and the minimum relevance score.
  4. From then on, every prediction for that capability retrieves first and answers with your documents in hand.

Next