Multiple indexes

Add each source separately

import { AgentBuilder } from '@anvia/core'
import { vectorFilter } from '@anvia/core/vector-store'

const agent = new AgentBuilder('support', model)
  .instructions(
    'Answer from public documentation and applicable support policy.',
  )
  .dynamicContext(publicDocsIndex, {
    topK: 4,
    threshold: 0.72,
    filter: vectorFilter.eq('published', true),
  })
  .dynamicContext(policyIndex, {
    topK: 2,
    threshold: 0.78,
    filter: vectorFilter.and(
      vectorFilter.eq('audience', 'support'),
      vectorFilter.eq('region', user.region),
    ),
  })
  .build()

Anvia searches every registered index before the model turn and appends the formatted matches. Each index keeps its own topK, threshold, filter, and formatter.

Budget the combined result

The maximum retrieved document count is the sum of each index's topK. In the example, the model can receive up to six documents per turn.

Set smaller limits for secondary sources. More results do not automatically produce a better answer; they can bury the strongest evidence and consume the model's context window.

Use separate indexes when boundaries differ

Separate indexes are useful when sources have different:

permissions or tenant scopes
embedding models or vector-store adapters
update schedules
document shapes and formatters
relevance thresholds

Use one index when the documents share the same lifecycle and access policy. Metadata filters are usually simpler than splitting an otherwise uniform corpus.

Keep source identity visible

Format each source with a recognizable ID and label:

const policyOptions = {
  topK: 2,
  threshold: 0.78,
  format(result) {
    return {
      id: `policy:${result.id}`,
      text: `Source: Internal policy\n\n${String(result.document)}`,
    }
  },
} satisfies Parameters<AgentBuilder['dynamicContext']>[1]

Clear source identity helps the model distinguish product documentation from internal policy and makes retrieval traces easier to inspect.

Tune sources independently

Test which index supplied each answer. If one source dominates with weak matches, raise its threshold or reduce its topK instead of changing every index. Keep permission tests separate for each source because every registered index is an independent path into model context.

Multiple indexes ​

Add each source separately ​

Budget the combined result ​

Use separate indexes when boundaries differ ​