Automatic retrieval

Use dynamic context when most prompts should receive a small set of relevant documents automatically.

Add an index

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

const agent = new AgentBuilder('docs-support', model)
  .instructions(
    'Answer with retrieved documentation when it is relevant.',
  )
  .dynamicContext(docsIndex, {
    topK: 4,
    threshold: 0.74,
    filter: vectorFilter.eq('published', true),
  })
  .build()

For every turn, Anvia extracts retrieval text from the current prompt, searches the index, applies its threshold and filter, and sends matching results as documents with the model request.

A later turn after a tool call can retrieve different documents because the runtime prompt has changed.

Format retrieved documents

Use format(...) when a stored object needs a concise, source-aware shape.

const policyContext = {
  topK: 3,
  threshold: 0.76,
  format(result) {
    return {
      id: `policy:${result.id}`,
      text: [
        `Title: ${result.metadata?.title ?? 'Untitled'}`,
        `Source: ${result.metadata?.source ?? 'unknown'}`,
        '',
        String(result.document),
      ].join('\n'),
    }
  },
} satisfies Parameters<AgentBuilder['dynamicContext']>[1]

Keep the formatted context focused. Retrieval should not become an unbounded corpus dump.

Tune with real prompts

Lower topK when extra context distracts the model. Raise threshold when weak matches appear. Improve chunking when answers span unrelated passages, and tighten filters when stale or unauthorized documents are eligible.