Anvia

Knowledges ​

Knowledge gives an agent access to a large document collection without sending the whole collection to the model. Anvia separates this into an ingestion path and a retrieval path:

text
Ingestion: source files -> documents -> embeddings -> vector store
Runtime:   user prompt -> filtered search -> relevant documents -> model

Ingestion normally runs in a script, worker, or deployment job. Runtime retrieval should only search an index that is already prepared.

1. Prepare searchable documents ​

For the common path, pass raw text documents to ingestVectorDocuments(). It applies shared deterministic chunking, embeds each document's chunks, and upserts the complete document groups with stable IDs:

ts
import { readFile } from 'node:fs/promises';
import type { TextDocument } from '@anvia/core/documents'
import {
  InMemoryVectorStore,
  ingestVectorDocuments,
} from '@anvia/core/vector-store'

const path = 'content/support/reset-links.md';
const text = await readFile(path, 'utf8');

const store = new InMemoryVectorStore<TextDocument>()

await ingestVectorDocuments({
  store,
  documents: [{
    id: path,
    text,
    metadata: { source: path, published: true },
  }],
  embeddingModel,
  chunking: {
    strategy: 'recursive',
    maxSize: 1_600,
    overlap: 200,
    separators: ['\n\n', '\n', '. ', ' '],
  },
})

The in-memory store is useful for learning and tests. Use a persistent vector-store adapter for production data.

2. Connect retrieval to an agent ​

Use createVectorContext() when the agent should retrieve relevant documents before each model turn:

ts
import { Agent, createVectorContext } from '@anvia/core'
import { vectorFilter } from '@anvia/core/vector-store'

const agent = new Agent({
  id: 'docs-support',
  model,
  instructions: 'Answer from the documentation. Say when the answer is not available.',
  context: [
    createVectorContext({
      store,
      model: embeddingModel,
      topK: 4,
      minScore: 0.72,
      filter: vectorFilter.eq('published', true),
    }),
  ],
})

const result = await agent.generate({
    prompt: 'How long does a reset link last?'
})

The index uses the current prompt as its search query. Only matching documents are added to the model request.

3. Choose the retrieval pattern ​

Use automatic retrieval when knowledge is useful for most prompts. Use a search tool when search is optional or the model may need to refine the query.

Use Knowledge GraphRAG when typed relationships and bounded traversal add evidence that independent document similarity cannot represent. Portable graph primitives work with both Neo4j and Memgraph.

Use static agent context for a small set of facts that should be present on every turn. Use an application tool for live records, permission checks, and actions.

Continue through the ingestion flow:

Built for Anvia.