GraphRAG is a retrieval-augmented generation (RAG) method for questions that require connections across many documents. Instead of retrieving only text passages that resemble a query, it builds an LLM-derived knowledge graph, organizes related entities into hierarchical communities, and prepares summaries that can support corpus-wide answers. That extra structure can improve thematic synthesis, but it also adds indexing work, model usage, latency, and tuning. GraphRAG is therefore a complement to vector RAG—not a universal replacement.
Why ordinary RAG can miss corpus-wide answers
Conventional RAG usually embeds document chunks, retrieves the top results by vector similarity, and gives those passages to a language model. This works well when the answer is stated in, or closely related to, a small number of passages.
As an Amazon Associate I earn from qualifying purchases.
It becomes less reliable when the question asks for aggregation or broad interpretation. Examples include “What are the main themes in this dataset?” or “Catch me up on the last two weeks of updates.” No single passage may contain the answer, and the wording of a global question may not resemble the wording scattered across the relevant documents. A top-k retriever can consequently return individually relevant snippets without providing a complete view of the collection.
Microsoft Research’s 2024 paper studied this class of global sensemaking questions on datasets in the one-million-token range. It reported substantial gains in answer comprehensiveness and diversity over a conventional RAG baseline, but that result is bounded to the tested task and scale; it is not a general accuracy guarantee for every workload.
What GraphRAG adds
GraphRAG precomputes a structured representation of the corpus. An indexing pipeline divides source material into TextUnits, uses an LLM to extract entities, relationships, and key claims, and stores those elements as a knowledge graph. A hierarchical community-detection process based on the Leiden technique groups closely connected entities. The system then generates summaries—often called community reports—from the bottom up.
At query time, these reports and graph-derived facts become model context. The original “From Local to Global” approach describes answering a broad question by producing partial responses from relevant community summaries and reducing them into a final response. For a question about a particular entity, GraphRAG can combine graph information with the underlying text chunks instead of relying on a summary alone.
The indexing pipeline
- Chunk the corpus: source documents are split into TextUnits that can be processed and retrieved.
- Extract structure: an LLM identifies entities, relationships, and important claims in those units.
- Build communities: the extracted graph is clustered hierarchically with Leiden-based community detection.
- Summarize communities: related entities and evidence are summarized from lower levels upward, creating reports at multiple granularities.
- Answer with derived context: a selected search mode supplies reports, graph facts, and/or source chunks to the answering model.
Because extraction and summarization are model-generated, the graph is not an infallible copy of the documents. Prompt design, model choice, source quality, and indexing configuration can affect what structure is created. Microsoft’s documentation recommends prompt tuning for a particular dataset rather than assuming the default prompts are optimal.
GraphRAG search modes compared
| Mode | Best fit | How it works | Main trade-off |
|---|---|---|---|
| Global Search | Corpus-wide themes, comparisons, aggregation, and holistic sensemaking | Uses community reports in a map-reduce process: generate partial answers from reports, then combine them into a response | Resource-intensive; answer detail, latency, and LLM usage depend on the hierarchy level used for reports |
| Local Search | Questions centered on a known person, product, organization, event, or other entity | Joins graph-derived entity information with relevant raw text chunks | More focused than global search, but still carries the cost and possible errors of graph extraction |
| DRIFT Search | Entity-level questions that may benefit from broader context and iterative investigation | Starts with relevant community reports, generates follow-up questions, and refines the answer through local search | Additional model calls and query steps can increase latency and resource use |
| Basic Search | Questions suited to ordinary top-k vector retrieval | Provides a rudimentary vector-RAG path for comparison with graph-based modes | Does not provide GraphRAG’s community-level synthesis |
When GraphRAG is a good fit
Use it for global sensemaking
Choose GraphRAG when users regularly ask for themes, trends, cross-document comparisons, timelines, or an overview of a large private collection. These questions require evidence distributed across the corpus, so precomputed communities and reports give the model an organized starting point.
Use local or DRIFT search for entity questions
If a user asks about one company, incident, project, or named individual, local search is usually the more direct mode. DRIFT is useful when that local question needs wider context and follow-up exploration rather than a single narrowly retrieved passage.
Keep basic vector retrieval for straightforward lookups
A policy question answered by one document, a search for an exact procedure, or another passage-level lookup may not justify graph indexing. Basic search also gives teams a practical baseline for deciding whether the added structure improves their own query mix.
Costs, limitations, and operational choices
Indexing is more involved
GraphRAG performs LLM extraction, graph construction, clustering, and report generation before users submit questions. That preprocessing consumes model inference, storage, and engineering time. Changes to the source corpus or prompts can require reindexing or regeneration of derived reports.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsGlobal queries can be expensive
Global Search evaluates community reports through a map-reduce workflow. The Microsoft GraphRAG documentation warns that lower-level reports can make responses more thorough while increasing response time and LLM resource use. It also states that “the quality of the global search’s response can be heavily influenced by the level of the community hierarchy chosen for sourcing community reports.” Selecting a higher or lower level is therefore a quality-versus-cost decision, not a fixed setting that is best for every corpus.
A graph does not remove hallucinations
The graph and reports are generated representations, not independently verified facts. Extraction can omit a relationship, merge entities incorrectly, or summarize away an important qualification. The answering model can still misread or overstate the supplied context. Retaining links to source TextUnits, inspecting representative communities, and evaluating answers against known questions are important safeguards.
Prompt tuning matters
Prompts control how entities, relationships, claims, and summaries are extracted. Terminology, document style, desired citation behavior, and the meaning of “community” vary by dataset. Tune and evaluate the indexing and query prompts with a representative question set instead of treating out-of-the-box behavior as final.
A practical decision framework
- Question scope: passage-level lookup favors basic vector search; an entity-centered investigation favors local or DRIFT; a corpus-wide synthesis favors global search.
- Evidence distribution: if the answer is likely spread across many documents, graph communities can organize the evidence before generation.
- Freshness: rapidly changing collections must absorb indexing delay and the cost of updating reports.
- Latency and budget: global and DRIFT modes make more model calls than a simple top-k query.
- Traceability: require a way to inspect source chunks and graph-derived claims when answers influence decisions.
- Evaluation: compare modes on the questions users actually ask, measuring completeness, diversity, factual grounding, latency, and resource consumption.
A hybrid deployment is often the sensible design: route simple lookups to vector retrieval, entity questions to local search, exploratory entity questions to DRIFT, and broad analytical requests to global search.
What the evidence supports—and what it does not
The Microsoft Research paper supports improvements in comprehensiveness and diversity for a defined class of global sensemaking questions, including datasets around one million tokens. It does not establish that GraphRAG always outperforms vector RAG, nor does it provide a universal cost multiplier or accuracy percentage.
Microsoft lists GraphRAG and LazyGraphRAG as available through Microsoft Discovery, an agentic scientific-research platform in Azure. Availability and product details can change, so that listing should be treated as an access route rather than proof that every GraphRAG deployment requires that platform.
Bottom line
GraphRAG addresses a specific weakness of conventional RAG: answering questions whose evidence and themes are distributed across an entire corpus. Its entity graph, hierarchical communities, and generated reports give global search a structured basis for synthesis, while local and DRIFT modes extend graph context to entity-focused questions. The price is additional preprocessing, model usage, tuning, and operational complexity. Match the mode to the question, keep vector retrieval where it is sufficient, and validate graph-derived answers against the underlying documents.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →
