FalkorDB GraphRAG SDK: Fast Knowledge Graph Generation for GenAI

GraphRAG-SDK has its own documentation at docs.falkordb.com/graphrag. This page is a short overview — the full guides, API reference, and runnable examples live there.

Build intelligent GraphRAG applications with FalkorDB and LLMs

GraphRAG-SDK builds knowledge graphs from your documents and tables and answers questions over them using graph-based retrieval-augmented generation, on FalkorDB. It ingests text, Markdown, and PDF with schema-guided entity extraction, and CSV and other delimited tables through a declared mapping, so a table row and a paragraph about the same thing become one node. It retrieves cited answers through a hybrid pipeline that combines vector search, full-text search, and relationship expansion, with optional text-to-Cypher for counts and aggregates. It merges duplicate entities across documents with an LLM judge that checks only what is new, and re-syncs individual documents and tables incrementally without rebuilding the graph. It holds the top score on GraphRAG-Bench across both the multi-document and single-document question sets, and works with any LLM or embedder reachable through LiteLLM.

Quick start

pip install "graphrag-sdk[litellm,pdf]"
docker run -p 6379:6379 falkordb/falkordb:v4.22.2

GraphRAG-SDK needs FalkorDB 4.20 or later (FalkorDB 6.0 is supported) and refuses an older server at its first query with FalkorDBVersionError. Keep the image tag: a bare falkordb/falkordb reuses whatever image Docker has cached, which may be older.

import asyncio
from graphrag_sdk import GraphRAG, ConnectionConfig, LiteLLM, LiteLLMEmbedder

async def main():
    async with GraphRAG(
        connection=ConnectionConfig(host="localhost", graph_name="my_graph"),
        llm=LiteLLM(model="openai/gpt-4o"),
        # 256 dimensions, to match GraphRAG's default embedding_dimension
        embedder=LiteLLMEmbedder(model="openai/text-embedding-3-small", dimensions=256),
    ) as rag:
        await rag.ingest("my_document.pdf")
        await rag.finalize()  # merge duplicate entities, embed, and index
        answer = await rag.completion("What is the main topic?")
        print(answer.answer)

asyncio.run(main())

Or skip the local instance and connect to FalkorDB Cloud.

Documentation

Guide	Description
Getting Started	Full tutorial from install to first query
Architecture	How the ingestion and retrieval pipeline works
Configuration	Connection, providers, and tuning reference
Graph Schema	The node types, edges, properties, and indexes in the graph, and how to declare your own ontology
Ingestion	Document loading, chunking, extraction, and resolution
Structured Ingestion	Load CSV and other tables through a declared mapping, as typed entities in the same graph
Retrieval	Hybrid retrieval, expansion, and reranking
Strategies	Swappable strategy interfaces and built-in implementations
Providers	LLM and embedder configuration
Storage	How nodes, embeddings, and indexes are stored, how `finalize()` merges duplicate entities, and the FalkorDB server requirements
Extraction	Entity and relation extraction internals
Incremental Updates	`update()`, `delete_document()`, and `apply_changes()` for CI-driven syncs
Ontology Discovery	Auto-draft a schema from a corpus
Ontology Evolution	Safely change a schema on a populated graph
Reliability and Grounding	Provenance, evidence trails, confidence thresholds, and abstention, mapped to the APIs
Reducing LLM Hallucinations	Grounded retrieval and an explicit abstention path for verifiable answers
Benchmark	Methodology, results, and reproduction instructions
Accuracy Benchmark	GraphRAG-SDK compared with vector RAG on GraphRAG-Bench, with full reproduction details
API Reference	Full API documentation

GraphRAG Server — the hosted, no-code option

If you’d rather not run the pipeline yourself, GraphRAG Server is the fully hosted version. Upload your files and it builds the knowledge graph for you, so you can ask questions in plain language and get accurate, cited answers — no code, no Cypher, and no graph expertise required.

GraphRAG Server: turn your documents into a knowledge graph you can question

How it works

Sign in at graphrag.falkordb.com with your Google account.
Upload your documents: PDF, plain text, or Markdown. GraphRAG Server reads them and builds your knowledge graph automatically, showing live progress as it goes.
Ask questions in plain language. You get an answer with citations to your sources, plus a small “explainability” graph showing the facts behind it.
Explore visually: browse the entities and relationships extracted from your documents in an interactive graph view.

Settings are there if you want them, but the defaults work out of the box.

Key features

Document upload: bring your own PDF, TXT, or Markdown files.
Natural-language Q&A: ask in plain English and get grounded answers.
Citations: every answer links back to the source documents it used.
Explainability subgraphs: see the entities and relationships behind each answer.
Interactive graph visualization: explore your knowledge as a graph you can click through.
AI-suggested questions: get starter questions tailored to your own content.
Embeddable chat widget: publish a finished graph and embed it on your own website or documentation so visitors can ask questions without leaving your site.

Why GraphRAG?

Vanilla RAG (Retrieval-Augmented Generation) finds text snippets that look similar to your question and hands them to an LLM. That works for simple lookups, but it struggles when the answer depends on how facts connect across different documents. GraphRAG first organizes your documents into a knowledge graph: the people, places, products, and concepts in your content, and the relationships between them. Asking questions against that graph gives you:

More accurate answers: it connects related facts across your documents instead of matching keywords in isolation.
Cited answers: every response points back to the source documents it came from, so you can trust and verify it.
Explainability: see the small graph of entities and relationships behind each answer, so you understand why you got it.
A view of your knowledge: explore everything the system learned from your documents as an interactive graph.

Resources:

GraphRAG-SDK documentation
GraphRAG-SDK GitHub repository
Runnable examples
Try GraphRAG Server — the hosted, no-code version
Understanding Ontologies and Knowledge Graphs

Frequently Asked Questions

What is GraphRAG-SDK?

GraphRAG-SDK builds knowledge graphs from your documents and tables on FalkorDB and answers questions over them using graph-based retrieval-augmented generation. It ingests text, Markdown, and PDF with schema-guided entity extraction, loads CSV and other tables through a declared mapping, and retrieves cited answers through a hybrid pipeline combining vector search, full-text search, and relationship expansion, with optional text-to-Cypher.

How do I get started with GraphRAG-SDK?

Install it with pip install "graphrag-sdk[litellm,pdf]", start FalkorDB 4.20 or later with docker run -p 6379:6379 falkordb/falkordb:v4.22.2, then create a GraphRAG instance with a ConnectionConfig, an LLM, and an embedder. Call ingest(), then finalize(), then completion() to ask your first question. The pdf extra is only needed for PDF ingestion — drop it if you are loading plain text, and add the markdown extra for Markdown files. You can also skip the local instance and point it at FalkorDB Cloud.

What is the difference between GraphRAG-SDK and GraphRAG Server?

GraphRAG-SDK is the self-hosted Python library you run and configure yourself. GraphRAG Server is the fully hosted, no-code version: you upload documents through the browser and it builds the graph for you, with no code, no Cypher, and no graph expertise needed.

Which LLMs and embedding models does GraphRAG-SDK support?

Any LLM or embedder reachable through LiteLLM, which covers OpenAI, Anthropic, Google, and locally hosted models. You select them with LiteLLM(model=...) and LiteLLMEmbedder(model=...).

What file types can GraphRAG-SDK ingest?

Plain text, Markdown, and PDF as documents. PDF ingestion needs the pdf extra and Markdown the markdown extra (pip install "graphrag-sdk[litellm,pdf,markdown]"). Tables in .csv, .tsv, .psv, or .tab files are read as records rather than prose: you declare what the columns mean with a TableMapping in the ontology, and each row becomes a typed entity with no LLM involved. If no mapping is declared, the SDK proposes one. See Structured Ingestion. GraphRAG Server accepts plain text, Markdown, and PDF through its upload interface.

Which FalkorDB version does GraphRAG-SDK need?

FalkorDB 4.20 or later, including FalkorDB 6.0. The SDK checks the server version once per connection and refuses an older server with FalkorDBVersionError, because earlier versions dropped vector-index entries on writes to embedded nodes, so search and deduplication missed entities. Upgrade the server before upgrading the SDK.

How does GraphRAG-SDK handle duplicate entities?

finalize() merges duplicates across documents. It checks only the entities added since the last run, compares each with its nearest names in the vector index, and asks an LLM judge to decide which candidates are the same entity. Every merge keeps all descriptions, records the other names as aliases, and moves every relationship to the surviving entity. An interrupted run is finished by the next one. See Storage and Incremental Updates.

Can GraphRAG-SDK answer counting and aggregate questions?

Yes, with text-to-Cypher turned on: GraphRAG(..., enable_cypher=True). It generates a Cypher query for questions no passage answers directly, such as counts and averages, and uses the column types a table mapping declares. It is off by default because it adds an LLM call per question (up to 3 when a generated query fails validation and is regenerated). You can also run your own read query with rag.query(cypher, params).

Do I have to rebuild the whole graph when a document changes?

No. GraphRAG-SDK re-syncs individual documents incrementally, and re-ingesting a table re-syncs it as a snapshot: deleted rows are removed and changed cells are updated. See Incremental Updates for update(), delete_document(), and apply_changes(), which are designed for CI-driven syncs, and drop_table() to remove a table and everything it wrote.

Why use GraphRAG instead of regular RAG?

Vanilla RAG retrieves text snippets that look similar to your question, which breaks down when an answer depends on how facts connect across documents. GraphRAG first organizes your content into a knowledge graph, so answers connect related facts across documents, come with citations back to the sources, and include an explainability subgraph showing why you got that answer.

How accurate is GraphRAG-SDK?

It holds the top score on GraphRAG-Bench across both the multi-document and single-document question sets. The Benchmark page documents the methodology and how to reproduce the results.

Where is the full GraphRAG-SDK documentation?

This page is a short overview. The complete guides, API reference, and runnable examples live at docs.falkordb.com/graphrag.

​Build intelligent GraphRAG applications with FalkorDB and LLMs

​Quick start

​Documentation

​GraphRAG Server — the hosted, no-code option

​How it works

​Key features

​Why GraphRAG?

​Frequently Asked Questions

Build intelligent GraphRAG applications with FalkorDB and LLMs

Quick start

Documentation

GraphRAG Server — the hosted, no-code option

How it works

Key features

Why GraphRAG?

Frequently Asked Questions