Best RAG Frameworks (2026): LlamaIndex, LangChain, Haystack and More

Quick picks

Licences, release dates and maintenance status were checked on GitHub on 6 October 2026; features come from each project’s own documentation.

Library, engine or platform: what you are choosing

“RAG framework” covers three different things, and mixing them up wastes evaluation time.

• Libraries such as LlamaIndex, LangChain, Haystack, DSPy and txtai give you components to assemble in your own code. You own the service, the API, the deployment and the operations.
• Engines and applications such as RAGFlow, Kotaemon and LightRAG’s server ship a running system with a UI and an API. You configure more than you code, and you accept their architecture.
• Platforms add identity, permissions, audit and administration across many use cases. Our page on enterprise RAG describes what that layer has to do.

A framework also sits on top of a vector database, which is a separate decision. Our list of vector databases for RAG covers that layer.

RAG frameworks compared

Licence, release and maintenance status

Capabilities

A dash in this table means we found no documentation of the capability on the pages we read; check the project before ruling it out.

Notes on each framework

LlamaIndex

LlamaIndex is the most integration-heavy option: core plus more than 300 packages for models, embeddings and vector stores. Its evaluation module is the most complete of the libraries here, with LLM-based correctness, semantic similarity, faithfulness, context relevancy, answer relevancy and guideline evaluators, ranking metrics such as hit rate and MRR for retrievers, synthetic question generation, and integrations with Ragas, DeepEval and others. Know the company context, though. In September 2026 the README added a note that the company’s primary focus has shifted to LlamaParse, its commercial parsing, extraction and indexing platform, while the framework remains available as an open toolkit. Releases continue on a regular cadence, so this is a signal to watch rather than an exit sign. For agents, LlamaIndex now builds on Agent Workflows, covered in our agent framework list.

LangChain and LangGraph

LangChain’s documentation names five building blocks for retrieval: document loaders, text splitters, embedding models, vector stores and retrievers. It describes three architectures: 2-step RAG, where retrieval always runs first; agentic RAG, where an agent decides when and how to retrieve; and a hybrid of the two with validation steps. LangGraph supplies what agentic RAG needs in production, namely durable execution that resumes after failures, human-in-the-loop interrupts and short- and long-term memory. Both packages reached 1.0 on 17 October 2025. Evaluation can run through LangSmith, a commercial service, or the MIT-licensed openevals package, whose RAG evaluators cover correctness, helpfulness, groundedness and retrieval relevance.

Haystack

Haystack, from deepset, builds RAG and agents as explicit pipelines of components with loops, branches and conditional routing, which makes every step inspectable. Version 3.0, released on 20 July 2026, gave the Agent hooks such as before_llm and before_tool for guardrails and human-in-the-loop checkpoints, plus first-class skills and async serving. Evaluation is built in: statistical evaluators for exact match, MRR, MAP, NDCG and recall, and model-based ones for context relevance, faithfulness and semantic answer similarity, with Ragas and DeepEval wrappers. Hayhooks serves a pipeline as a REST API, an MCP server or an OpenAI-compatible chat endpoint. One setting to change for private deployments: Haystack sends anonymous component-usage telemetry by default, which HAYSTACK_TELEMETRY_ENABLED=False switches off.

RAGFlow

RAGFlow is a self-hosted RAG engine with a web interface, not a library. Its strength is parsing: the DeepDoc module handles layout analysis, OCR and table recognition across Word, slides, Excel, text, images, scans and web pages, and chunking is template-based with a visual view so people can correct chunks by hand. Retrieval uses multiple recall paths with fused re-ranking, and August 2026 releases added agentic RAG with four thinking modes and “knowledge compilation” into wikis, graphs, trees and timelines. Answers carry traceable citations. The 1.0.0-rc1 release of 29 September 2026 moves the server to Go. The README recommends at least 4 CPU cores, 16 GB of RAM and 50 GB of disk, with Elasticsearch or Infinity as the document engine. Our guide to OCR for scanned documents explains why parsing quality decides so much downstream.

DSPy

DSPy, from Stanford NLP, treats a RAG system as a program whose prompts and weights are optimised against a metric. Its RAG tutorial retrieves with dspy.retrievers.Embeddings, scores answers with dspy.Evaluate and a SemanticF1 metric, and compiles the program with the MIPROv2 optimizer, which in that tutorial lifted semantic F1 on the development set from about 42% to 61%. That is the project’s own example, not a general benchmark. DSPy brings no connectors or ingestion layer, so teams usually pair it with another framework or their own loaders, and a good evaluation set is a prerequisite.

txtai

txtai, from NeuML, centres on an embeddings database that combines sparse and dense vector indexes, graph networks and a relational store. Its RAG pipeline draws context from that database and generates with Hugging Face Transformers, llama.cpp or LiteLLM backends, and its examples cover citations and graph path traversal for multi-hop questions. It also runs agents, exposes web and MCP APIs, and installs with pip or Docker. It suits small teams that want one dependency instead of five.

Kotaemon and LightRAG

Kotaemon is a Gradio-based document chat interface with a hybrid full-text and vector retriever, reranking, question decomposition, ReAct and ReWOO agents, and citations shown with highlights in an in-browser PDF viewer. Docker images include a variant bundled with Ollama for fully local use. Activity has slowed, with the latest release on 31 May 2026. LightRAG, from the HKUDS group, adds a knowledge graph to vector retrieval, stores data in PostgreSQL, MongoDB, Neo4j or OpenSearch, and integrated RAGAS evaluation and Langfuse tracing in late 2025. When relationships across documents matter more than single passages, read our explainer on knowledge graph RAG first.

Archived and slowing projects

Four names still appear in older comparisons. Weaviate’s Verba and TrueFoundry’s Cognita are archived on GitHub. Microsoft’s GraphRAG README says the project is largely in maintenance mode and will not accept new features, and warns that its indexing can be expensive. R2R is not archived, but its latest release is from June 2025 and its last commit from November 2025. None is a sound base for a new production system.

How to choose a RAG framework

1. Decide library or engine. If you have engineers who will own a service, pick a library. If you need working document Q&A next month, an engine such as RAGFlow gets you further faster.
2. Test parsing on your worst documents. Scanned contracts, nested tables and slide decks break more pipelines than retrieval settings do. Our guide to chunking enterprise documents covers what to check after parsing.
3. Check the retrieval you need is first-class. Hybrid keyword and vector search, metadata filters and reranking should be configuration, not custom code.
4. Decide whether retrieval must be agentic. Multi-hop questions and tool use favour LangGraph, Haystack’s agent or LlamaIndex workflows; most FAQ-style assistants do not need it. Our comparison of agentic and traditional RAG sets out when it pays.
5. Wire evaluation in before features. Pick a framework whose evaluators you will actually run in CI.
6. Check maintenance, not stars. Look at the latest release, the last commit and any status note in the README.
7. Audit the defaults. Telemetry, hosted tracing and cloud parsing services can send data out of a private deployment unless you switch them off.

Evaluating a RAG pipeline before production

Whichever framework you choose, accuracy is measured outside it. Build a labelled set of real user questions and the passages that answer them, and score retrieval and generation separately:

Run the set after every change to chunking, embeddings, retrieval settings, prompts or the model, and keep the results with the release. Our RAG accuracy evaluation framework describes the method and the thresholds regulated teams use.

How VDF AI fits

Frameworks give you the building blocks; VDF AI ships the retrieval layer as part of the product. In VDF AI Chat, connectors, chunking, embedding, a local vector store and a permission filter form one layer, and retrieval is filtered by the reader’s permissions at query time. Answers cite their sources, and the governance layer writes append-only logs for every turn, applies retention policy and exports to your SIEM.

Teams that want more control can build vector indexes in VDF AI Data with their own chunk size, overlap and embedding model, and restrict sensitive knowledge sources to named people and agents. In VDF AI Networks, a Data Source node carries retrieval settings, embedding models and query templates into multi-step workflows, so retrieval becomes one governed step in a larger workflow. Everything runs on-premises, in a private cloud or air-gapped.

Sources

Verified 6 October 2026.

• LlamaIndex repository and README
• LlamaIndex evaluation guide
• LlamaAgents and Agent Workflows repository
• LangChain repository
• LangChain retrieval documentation
• LangGraph repository
• openevals repository
• Haystack repository
• Haystack 3.0.0 release notes
• Haystack evaluation
• Haystack telemetry
• Hayhooks repository
• RAGFlow repository
• RAGFlow releases
• DSPy repository
• DSPy RAG tutorial
• txtai repository
• txtai RAG pipeline
• Kotaemon repository
• LightRAG repository
• Microsoft GraphRAG repository
• R2R repository
• Verba repository (archived)
• Cognita repository (archived)