# Vector RAG vs GraphRAG vs LightRAG: Which to Use (2026)

> Compare vector RAG, Microsoft GraphRAG and LightRAG: what each indexes, which questions each answers, what each costs to run, and how to choose.

Source: https://www.super-cat.tech/guides/vector-rag-vs-graphrag-vs-lightrag  
Last modified: 2026-10-25

[Home](https://www.super-cat.tech/)/ [Guides](https://www.super-cat.tech/guides)

# Vector RAG vs GraphRAG vs LightRAG: which should you use?

Start with vector RAG: it answers questions that sit in a few passages and is the cheapest to build and run. Add GraphRAG when users ask corpus-wide questions such as "what are the main themes", and consider LightRAG when you want a graph without community summaries. As of 2026, Super Cat Technology adds a graph only where vector search alone fails.
[Cat Yung](https://www.super-cat.tech/cat-yung), Founder, Super Cat Technology Published 25 October 2026

In short

- The deciding factor is the question, not the tool: single-passage lookups, cross-document questions and whole-corpus questions need different retrieval.

- Microsoft’s GraphRAG paper reports that standard RAG fails on global questions about an entire corpus; that is the case GraphRAG was built for.

- LightRAG puts a graph into indexing and retrieval without community reports, and its documentation covers incremental updates and document deletion.

- Both graph methods run an LLM over your documents to extract entities, so they cost more to index than vector RAG and need their own evaluation.

## What is the difference between vector RAG, GraphRAG and LightRAG?

All three retrieve from your documents before an LLM writes the answer. They differ in what they index. Vector RAG indexes text chunks as embeddings. GraphRAG and LightRAG also have an LLM extract entities and relationships into a knowledge graph, then retrieve through that graph.
Animation in three steps. Vector RAG: the question is turned into a vector, the closest chunks are picked, and the LLM answers from them. GraphRAG: an LLM extracts entities and relations, groups them into communities, and summarises each community. LightRAG: broad concepts and specific facts both match the question, and links join them into one answer.Vector RAG

- 1 The question is turned into a vector

- 2 The closest text chunks are picked

- 3 The LLM answers from those chunks

GraphRAG

- 1 An LLM extracts entities and relations

- 2 Entities are grouped into communities

- 3 Each community is summarised; the summaries answer whole-corpus questions

LightRAG

- 1 Broad concepts match the theme of the question

- 2 Specific facts match its details

- 3 Links join both levels into one answer

| Method | What it indexes | Suits | What to watch |
| --- | --- | --- | --- |
| Vector RAG | Text chunks as embeddings; retrieves the top-k most similar | Questions answered by one or a few passages | Fails on corpus-wide questions; flat chunks can return fragmented answers |
| GraphRAG (Microsoft) | A knowledge graph, a community hierarchy and a summary per community | Questions about the whole corpus, such as main themes | An LLM pass over the full corpus at indexing; the docs strongly recommend prompt tuning |
| LightRAG (HKU) | A knowledge graph plus vectors, retrieved at two levels: specific facts and broad concepts | Cross-document questions where you want a lighter graph and easier updates | LLM extraction on every chunk; default storage is not for production |

## When is vector RAG enough?

Vector RAG is enough when the answer lives in one passage or a handful of them: a rule number, a clause, a figure in a table, a procedure. Most enterprise question-answering sits here, and it is the right baseline to build and measure first.
Super Cat Technology’s [on-premise RAG system for a Big 4 firm](https://www.super-cat.tech/portfolio/enterprise-rag-big-4) is built on this layer. It blends keyword and semantic search, about 30 to 70, because rule numbers and names are what embeddings alone miss, and it returns the source paragraph with every answer. Most of the quality came from parsing and retrieval tuning, not from a graph.

## When do you need GraphRAG?

You need GraphRAG when users ask questions about the corpus as a whole. The GraphRAG paper by Edge and colleagues, revised in February 2025, states that RAG fails on global questions such as “What are the main themes in the dataset?”, because they are summarisation tasks, not retrieval tasks.
GraphRAG answers them by extracting a knowledge graph, building a community hierarchy and generating a summary for each community. Its documentation lists four query modes: global search over community summaries, local search around specific entities, DRIFT search, and basic vector search for questions that plain RAG handles well. Expect to tune its prompts to your data.

## When is LightRAG the better graph option?

LightRAG fits when you want graph-based retrieval across documents but not the community-report machinery. Its paper by Guo and colleagues, revised in April 2025, describes graph structures in indexing and a dual-level retrieval system covering low-level facts and high-level concepts.
Its maintainers state that it avoids community reports, which reduces LLM calls, and that it supports incremental updates and selective document deletion. For graph storage it supports Neo4j and Memgraph, and its documentation recommends PostgreSQL as the single production backend. The default in-memory storage is for testing only.

## What can go wrong with graph-based RAG?

A graph adds capability and adds failure modes. Plan for these before you commit:

- **Indexing cost.** Both methods run an LLM over your documents to extract entities and relationships. LightRAG’s documentation says it needs more capable models than traditional RAG for extraction and for answering from long, noisy context.

- **Extraction quality.** The graph is only as good as the entities the LLM pulls out. GraphRAG’s documentation says out-of-the-box results may not be the best and recommends prompt tuning.

- **Locked-in embeddings.** LightRAG’s documentation says the embedding model must be chosen before indexing, and changing it means re-embedding everything, with no re-embedding tool.

- **Bad parsing stays bad.** A graph built from a table split mid-row or an unread scan inherits the error. Layout-aware parsing and OCR come before any retrieval method.

## How do you choose between them?

Choose by the questions your users actually ask. These examples show which method each kind of question points to:

| Example question | Kind | Start with |
| --- | --- | --- |
| What does clause 4.2 of the policy say? | Single passage | Vector RAG with keyword search |
| Which suppliers appear in both the contract and the audit report? | Cross-document | LightRAG, or GraphRAG local search |
| What are the main risks across all 500 filings? | Whole corpus | GraphRAG global search |

Then work in this order:

- Build a vector and keyword baseline with citations, and a labelled evaluation set.

- Sort the questions it fails on: single-passage, cross-document, or whole-corpus.

- Add a graph only for the cross-document or whole-corpus failures, and re-run the evaluation.

- Keep the source citation on every answer, whichever method retrieved it.

Where the data is confidential, all of this can run on your own infrastructure. Super Cat’s [private on-premise platform for family offices](https://www.super-cat.tech/portfolio/super-chain-family-office) is retrieval over research and holdings with every answer cited. For the build-versus-buy decision, see [Enterprise RAG: build vs buy](https://www.super-cat.tech/guides/enterprise-rag-hong-kong-build-vs-buy).

## Frequently asked questions

### What is GraphRAG?

GraphRAG is Microsoft Research’s retrieval method that extracts a knowledge graph from your documents, builds a community hierarchy and summarises each community. It is designed for global questions about a whole corpus, which the GraphRAG paper says standard RAG fails on.

### Is LightRAG better than GraphRAG?

Neither is better in general. LightRAG’s maintainers claim lower indexing and query cost because it skips community reports. GraphRAG is built for whole-corpus summarisation questions. Test both on your own questions with a labelled evaluation set before choosing.

### Do I need a graph database such as Neo4j for graph RAG?

Not to start. LightRAG’s default storage is in-memory and meant for testing only. For production its documentation recommends PostgreSQL as a single backend, and it also supports Neo4j and Memgraph for graph storage, so a graph database is one option, not a requirement.

### Can graph RAG run on our own servers?

Yes for LightRAG. Its documentation describes local deployment of embedding, reranking and storage through Docker, and names Qwen3-30B-A3B-Instruct as a reasonable minimum local extraction model. Super Cat Technology builds on-premise retrieval for confidential data.

## Sources

- [Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv 2005.11401)](https://arxiv.org/abs/2005.11401) (May 2020)

- [Edge et al., From Local to Global: A Graph RAG Approach to Query-Focused Summarization (arXiv 2404.16130)](https://arxiv.org/abs/2404.16130) (revised February 2025)

- [Microsoft, GraphRAG documentation](https://microsoft.github.io/graphrag/) (checked October 2026)

- [Guo et al., LightRAG: Simple and Fast Retrieval-Augmented Generation (arXiv 2410.05779)](https://arxiv.org/abs/2410.05779) (revised April 2025)

- [HKUDS, LightRAG repository and documentation](https://github.com/HKUDS/LightRAG) (checked October 2026)

Enterprise RAG: build vs buy →
When a platform is enough and when to build.
Hong Kong RAG solution →
The service and the systems shipped.
Big 4 RAG case study →
Hybrid retrieval with a source paragraph on every answer.
About Cat Yung →
The founder who leads the RAG work.

## About the author

[Cat Yung](https://www.super-cat.tech/cat-yung) is the founder of Super Cat Technology, an AI agent engineering team in Hong Kong and London, and works as an AI expert, data scientist and fractional CTO. Over a decade in production AI, NLP and machine learning; Expert-Vetted top 1% on Upwork with a 100% Job Success Score (source: Upwork profile, October 2026).

First-Sprint Data Scientist

## A vetted engineer embedded in your team, a shipped deliverable by end of sprint 1.

Pick a time below and you will speak with me directly. Bring the use case, the data and the constraints, and I will give you a straight view on feasibility, effort and the first thing to ship.
[Book a call](https://calendly.com/super-chain/catyung?utm_source=super-cat.tech&utm_campaign=first-sprint-data-scientist&utm_content=guides%2Fvector-rag-vs-graphrag-vs-lightrag)Cat Yung
Hong Kong & London time zones

Book a call · 30 min with Cat Yung