Dynamic context

The Dynamic Context view browses document chunks exposed by the vector indexes registered on Studio agents. It helps you verify index contents and metadata before debugging semantic selection in an agent run.

Open http://localhost:4021/ui/knowledge/dynamic-context.

Register an index

const agent = new AgentBuilder('docs-support', model)
  .instructions('Answer from relevant documentation.')
  .dynamicContext(docsIndex, {
    topK: 4,
    threshold: 0.72,
    filter: vectorFilter.eq('product', 'platform'),
    format: (result) => ({
      id: result.id,
      text: result.document.text,
      additionalProps: {
        product: result.document.product,
        score: result.score,
      },
    }),
  })
  .build()

These options control runtime retrieval:

Option	Effect during an agent turn
topK	Maximum number of matching chunks selected for a generation.
threshold	Rejects matches below the configured similarity score.
filter	Restricts the eligible index records before selection.
format	Converts a search result into the document sent to the model.

Studio does not perform an arbitrary similarity search from this browser. It uses the optional index inspection contract to page through source items, applying the registration's filter. Runtime topK and threshold still govern prompt-time retrieval.

Browse source items

Choose a source when the selected agent has several dynamic context registrations. Each source is identified by its registration order. Browse the returned items to inspect:

• the stable vector record ID;
• document text, when the stored document exposes it;
• the raw JSON document for other shapes;
• stored metadata returned by the index inspector.

Use Load more to follow the index cursor. When the inspector reports a total, Studio also uses it for the source item count.

When chunks are not browseable

Search and inspection are separate index capabilities. A custom VectorSearchIndex can support runtime search without implementing inspect(...). In that case the agent can retrieve normally, but Studio reports that the source does not expose browseable chunks.

This does not mean the index is empty or broken. Add an inspection adapter if browsing is useful, or validate actual runtime selection through Retrieval evidence.

Diagnose retrieval in the right order

Symptom	Check first
Expected chunk is absent from the browser	Ingestion, chunking, upsert, registration filter, and selected source
Chunk exists but does not reach the generation	Query text, embedding model, threshold, topK, and filter
Correct chunk reaches the generation but answer is wrong	Instructions, surrounding messages, document formatting, and model output

Studio is intentionally not an ingestion console. Build and update indexes with the patterns in Add dynamic context, then restart or refresh Studio to inspect the registered source.