Graph RAG

Build a graph-augmented RAG system with document extraction, facts, entities, and weighted search.

Install the SDK

Cloud mode only needs the TypeGraph SDK. Self-hosted mode uses the pgvector adapter, an embedder, and an LLM. Bridge constructors are not part of the public setup path.

terminal

npm install @typegraph-ai/sdk

Initialize With an LLM

graphExtraction: true runs the configured extractor. In Cloud, TypeGraph manages extraction infrastructure. In self-hosted mode, pass vectorStore, embedding, optional searchEmbedding, and llm.

Initialize TypeGraph with graph support

import { TenantId, typegraphInit } from '@typegraph-ai/sdk'

const typegraph = await typegraphInit({
  apiKey: process.env.TYPEGRAPH_API_KEY!,
  tenantId: TenantId('tenant_acme'),
})

Create a Graph-Enabled Bucket

Buckets are named namespaces for documents and event artifacts. Enable graph extraction at the bucket level when most writes to that bucket should produce entities and facts.

Create bucket example

const bucket = await typegraph.bucket.create({
  id: 'research',
  name: 'Research',
  description: 'Research papers and technical notes',
  indexDefaults: {
    graphExtraction: true,
    deduplicateBy: ['content'],
  },
})

Ingest Documents With Graph Extraction

Document ingestion produces chunks and embeddings. When graph extraction is enabled, TypeGraph also extracts entities, facts, and provenance links from the document content.

Document ingest example

const result = await typegraph.document.ingest(
  [{
    id: 'paper:constitutional-ai',
    name: 'Constitutional AI Overview',
    description: 'Research overview covering Constitutional AI and related organizations.',
    content: 'Anthropic, founded by Dario and Daniela Amodei, focuses on AI safety research. Their work on Constitutional AI builds on reinforcement learning from human feedback...',
    metadata: { kind: 'research-note' },
  }],
  {
    bucketId: bucket.id,
    graphExtraction: true,
  },
)

Search With Graph Weighting

resources controls what can be returned. weights.graph controls how strongly graph evidence influences ranking. Use explain: true while tuning graph behavior.

Graph search example

const response = await typegraph.search('Who founded Anthropic?', {
  resources: ['documents', 'entities', 'facts'],
  weights: { semantic: 1, bm25: 0.5, graph: 0.9, recency: 0.2 },
  limit: 5,
  promptBuilder: {
    format: 'xml',
    sections: ['chunks', 'facts', 'entities'],
  },
  explain: true,
})

console.log(response.hits)
console.log(response.prompt)
console.log(response.explanation?.graphTrace)

Graph-Augmented RAG Pipeline

Use promptBuilder to produce the LLM-ready response.prompt. The graph contributes facts and entity context, while document chunks remain the grounded text evidence.

Graph RAG pipeline example

import { generateText } from 'ai'
import { openai } from '@ai-sdk/openai'

const question = 'What is Constitutional AI and who developed it?'

const response = await typegraph.search(question, {
  resources: ['documents', 'entities', 'facts'],
  weights: { semantic: 0.8, bm25: 0.4, graph: 1, recency: 0.2 },
  limit: 10,
  promptBuilder: {
    format: 'xml',
    sections: ['chunks', 'facts', 'entities'],
    maxTotalTokens: 6000,
  },
})

const { text } = await generateText({
  model: openai('gpt-5.4-mini'),
  prompt: `Answer the question using only the context below.

Context:
${response.prompt}

Question: ${question}`,
})

console.log(text)

Back to All Guides

const bucket = await typegraph.bucket.create({ id: 'research', name: 'Research', description: 'Research papers and technical notes', indexDefaults: { graphExtraction: true, deduplicateBy: ['content'], }, })

const result = await typegraph.document.ingest( [{ id: 'paper:constitutional-ai', name: 'Constitutional AI Overview', description: 'Research overview covering Constitutional AI and related organizations.', content: 'Anthropic, founded by Dario and Daniela Amodei, focuses on AI safety research. Their work on Constitutional AI builds on reinforcement learning from human feedback...', metadata: { kind: 'research-note' }, }], { bucketId: bucket.id, graphExtraction: true, }, )

const response = await typegraph.search('Who founded Anthropic?', { resources: ['documents', 'entities', 'facts'], weights: { semantic: 1, bm25: 0.5, graph: 0.9, recency: 0.2 }, limit: 5, promptBuilder: { format: 'xml', sections: ['chunks', 'facts', 'entities'], }, explain: true, }) console.log(response.hits) console.log(response.prompt) console.log(response.explanation?.graphTrace)

import { generateText } from 'ai' import { openai } from '@ai-sdk/openai' const question = 'What is Constitutional AI and who developed it?' const response = await typegraph.search(question, { resources: ['documents', 'entities', 'facts'], weights: { semantic: 0.8, bm25: 0.4, graph: 1, recency: 0.2 }, limit: 10, promptBuilder: { format: 'xml', sections: ['chunks', 'facts', 'entities'], maxTotalTokens: 6000, }, }) const { text } = await generateText({ model: openai('gpt-5.4-mini'), prompt: `Answer the question using only the context below. Context: ${response.prompt} Question: ${question}`, }) console.log(text)