The GraphRAG market is easier to compare when each product is described by what it consumes and what it produces. A tool that derives semantic definitions, a loader that writes business facts, and an application that answers questions all contribute different outputs—even when all three are described as knowledge-graph software.
This directory uses one alphabetical comparison matrix across frameworks, enterprise products, graph engines, and supporting tools. It contains 37 documented offerings, with primary-source links on their names.
Research snapshot: 5 November 2026. Descriptions are research summaries rather than tested integrations or performance scores. “GraphRAG” refers to the broader graph-assisted retrieval family; Microsoft GraphRAG is one particular implementation.
| Output | What it means here | Concrete example |
|---|---|---|
| Semantic model | Definitions of concepts, attributes, relationship types, identifiers, and source mappings | What a device is, what INSTALLED_AT means, and which source fields identify it |
| Ontology | Formal domain definitions and potentially logical axioms | Device classes, property meanings, and inference semantics |
| Populated knowledge graph | Identifiable entities, facts/assertions, relationships, and evidence | Device D-18 is installed at site S-9, according to record R-4 |
| Retrieved context | Evidence selected for a question | Relevant paths, records, passages, or summaries |
| Generated response | The answer composed from available context | A device-impact explanation with supporting references |
These are functional distinctions, not requirements for separate storage. Semantic definitions can themselves form a graph, and a knowledge graph can include both ontology and instance data. Producing a semantic model alone does not populate business facts. W3C RDF Schema, W3C OWL primer
Read product form alongside output. Database engines, modeling tools, retrieval libraries, and evaluation services are complementary roles rather than equivalent alternatives. Inputs are representative of the documented approach, not exhaustive connector lists.
| Tool / primary evidence | Product form | Starting input | Principal output | Dependency / scope note |
|---|---|---|---|---|
| Amazon Bedrock Knowledge Bases GraphRAG | Managed service | Documents in supported sources | Constructed graph and graph-enhanced answer context | Uses Neptune Analytics; inspect S3 and customization limits |
| Amazon Neptune Database | Graph service family | Graph data | Stored/queryable graph and analytical access | Database and Analytics are distinct services |
| ArangoDB | Multi-model platform | Graph/document/vector data | Queryable connected data and search results | Distinguish database and broader platform suite |
| AWS GraphRAG Toolkit | Developer toolkit | Documents or an existing graph | Lexical graphs or graph-grounded retrieval | Component, graph store, vector store, and model choices |
| Azure Cosmos DB for Apache Gremlin | Graph service | Graph data | Gremlin traversal results | API compatibility and partitioning matter |
| cognee | AI memory framework | Knowledge inputs | Persistent context and retrieval | Inspect configured graph/vector backends |
| DeepEval | Supporting evaluation | Test cases and application outputs | Quality metrics and test results | Define graph-specific assertions separately |
| Docling | Supporting parser | Documents | Structured document content and chunks | Parsing does not establish domain meaning or truth |
| FalkorDB | Graph database | Nodes and edges | Stored graph and traversal results | Runtime for its SDK/Server ecosystem |
| FalkorDB GraphRAG-SDK | Application SDK | Structured/unstructured inputs | Ontology proposals, populated graphs, and retrieval | Separate SDK, Server, and database responsibilities |
| fast-graphrag | Retrieval framework | Domain-guided document knowledge | Graph exploration and answer context | Personalized PageRank; distinct from Microsoft's fast method |
| Flow.BI | Semantic-model discovery | Metadata and profiled enterprise data | Concept definitions and source mappings | Separate ekg/doc-ekg loaders populate and enrich the graph |
| Google Spanner Graph | Graph/relational engine | Relational records and graph schema | Queryable property graph | Edition, dialect, and integration configuration apply |
| Graphiti | Temporal memory framework | Events, episodes, and changing facts | Temporal knowledge graph and retrieved context | Define current versus historical query behavior |
| Graphwise GraphRAG | Semantic application platform | Semantic context, graphs, and documents | Graph-grounded conversational responses | Graph Modeling and GraphDB contribute separate roles |
| HippoRAG | Research framework | External document knowledge | Graph-assisted retrieval using PageRank | Reproduce the selected project generation |
| KAG | Domain reasoning framework | Domain knowledge and schema | Aligned graph knowledge and reasoning/retrieval | OpenSPG ecosystem; assess schema and language coverage |
| LangChain graph integrations | Application framework components | Existing graph or constructed knowledge | Graph query access and composed workflows | Application defines extraction, retrieval, and access rules |
| Langfuse | Supporting observability | Application traces and evaluation data | Runtime, prompt, cost, and quality observations | Instrument graph queries and evidence selection |
| LightRAG | Graph/vector RAG framework | Documents | Entity/relationship index and retrieved context | Query-mode semantics differ from other implementations |
| LlamaIndex PropertyGraphIndex | Developer framework | Documents or property graph data | Extracted paths, graph indexes, and query results | Selected extractors, graph store, and retrievers determine behavior |
| Memgraph | Graph database ecosystem | Connected records | Graph storage and graph-backed queries | Check the specific GraphRAG integration |
| metaphactory | Semantic application platform | Domain models and enterprise knowledge | Semantic models and conversational graph applications | NL-to-SPARQL and unstructured RAG services differ |
| Microsoft GraphRAG | Research pipeline | Text corpus | Extracted graph, community reports, and search context | Repository states maintenance mode; multiple query modes |
| nano-graphrag | Compact framework | Text corpus | GraphRAG indexes and retrieval | Assess storage choices and maintenance suitability |
| Neo4j / Aura | Graph database family | Property graph data | Stored graph and Cypher query results | Edition, version, schema, and indexes matter |
| Neo4j GraphRAG for Python | GraphRAG library | Neo4j graph and/or construction inputs | Built graph, retrieved evidence, and generated responses | Requires supported Neo4j and configured models |
| Neo4j LLM Knowledge Graph Builder | Graph-building application | Unstructured documents | Entity/relationship graph linked to source chunks | Inspect extraction schema, identity merging, and evidence |
| Phoenix | Supporting observability | Application traces and test outputs | Evaluation results and runtime traces | General AI tooling; add graph-specific instrumentation |
| Protégé / WebProtégé | Ontology editor | Domain vocabulary and axioms | Formal ontology definitions | Produces semantics; not an automatic instance loader |
| PuppyGraph | Graph query engine | Relational data and mappings | Graph-query access over existing records | Semantic discovery and answer generation remain separate |
| Ragas | Supporting evaluation | Retrieved context and responses | Retrieval and answer metrics | Pair generic scores with known graph facts |
| Senzing | Supporting identity resolution | Fragmented entity records | Resolved identities and relationships | Entity resolution differs from defining concept types |
| Stardog Voicebox | Conversational application | Enterprise graph context and documents | Graph queries, analysis, and attributed responses | Verify which answer path supplied the evidence |
| TigerGraph GraphRAG | Enterprise graph/AI offering | Connected operational data | Graph/vector retrieval and application context | Specify database, integration, and application scope |
| TopBraid EDG | Semantic governance platform | Models, vocabulary, and content | Semantic assets and GraphRAG-enabling access | Check release, ingestion, and answering components |
| Unstructured | Supporting parser/extractor | Documents | Parsed content and structured fields | Retain source references through downstream ingestion |
| Market view | The distinction to inspect | Examples in the matrix |
|---|---|---|
| Knowledge foundations | Semantic definitions versus populated entity facts | Flow.BI, Protégé, semantic platforms, graph builders |
| Retrieval implementations | Query generation, traversal, graph exploration, or community search | Neo4j GraphRAG, fast-graphrag, HippoRAG, Microsoft GraphRAG |
| Application delivery | Developer components versus packaged or managed services | LlamaIndex, LangChain, AWS Toolkit, Bedrock Knowledge Bases, enterprise assistants |
| Production infrastructure | Stored graph, query access, preparation, and measurable quality | Graph engines, parsers, tracing, and evaluation tools |
These are reading aids, not exclusive categories. A vendor can span several outputs without every component having the same role.
| Names that overlap | Actual distinction |
|---|---|
| Microsoft GraphRAG and the wider GraphRAG market | One research implementation within a broader architectural family |
| Microsoft's FastGraphRAG and Circlemind fast-graphrag | An indexing method versus an independent project |
| AWS GraphRAG Toolkit and Bedrock Knowledge Bases GraphRAG | Open-source components versus a managed service |
| Neo4j database, GraphRAG library, and LLM Knowledge Graph Builder | Runtime, developer package, and construction application |
| FalkorDB database, GraphRAG-SDK, and GraphRAG Server | Engine, programmable toolkit, and packaged delivery |
| Flow.BI, flowbi/ekg, and flowbi/doc-ekg | Semantic derivation, structured graph loading, and document enrichment |
Graphiti and cognee emphasize persistent context and memory. They should be assessed on their own temporal and retrieval behavior rather than assumed to implement a community-summary pipeline. An orchestration graph describing agent steps is also different from a graph of domain knowledge used as evidence.
Two tools may both name a graph database while expecting different schemas, source-chunk links, embeddings, or indexes. A graph engine may also expose relational data without copying all records into a separate store. In either case, define the data and evidence contracts before claiming the components form a working stack.
For example, an equipment-impact query might require device identities, installation records, component versions, and manual passages. Test each connection and its provenance. A relevant paragraph alone does not establish which devices are affected; a relationship path alone may not explain the required action.
The source links support each product's described role. They do not establish universal accuracy, complete connector coverage, or a guarantee that graph-backed answers are grounded. Compare semantic meaning, identity, evidence paths, access control, freshness, and answer support on the same workload.
Use release-specific documentation for licensing, lifecycle, service regions, and dependencies. GraphDB and Stardog also supply graph storage within their wider ecosystems. This directory is broad public-tool coverage rather than a registry of every private accelerator.