We use cookies. Privacy choices

    Skip to content

    Live webinar Nov 5: save your seat

    Technical brief · 2026

    Fluree Reference Architecture

    What Fluree is, how its parts connect, what data moves and what stays put, and how it’s governed, deployed and licensed. Written for architects and data leaders.

    Updated Five-minute read

    flur.ee/resources/fluree-reference-architecture

    The short answer

    One governed graph between your data and your AI.

    What Fluree is and where it sits, in one paragraph.

    Fluree AI is the context layer for enterprise AI: one governed knowledge graph between the data you already run and the agents and apps that ask questions of it. It sits beside your warehouse or lakehouse, your catalog and your BI tools, and under your agents and apps. Nothing gets ripped out. Tables are read where they live, documents become facts with their evidence, and every answer passes the same policy check whether a person or an agent asked, with a history you can replay.

    • Apps and copilots

      Answers now cite the fact they came from.

      Keep
    • Agent frameworks

      One governed layer over MCP, instead of a retriever per agent.

      Keep
    • Vector search and RAG

      Search also runs inside the graph, joined to relationships and filtered by policy.

      Keep
    • BI and semantic layers

      Metrics keep their definitions. Entities and policy come from the graph.

      Keep
    • The enterprise AI context layerFluree

      Resolved entities, a governed ontology, policy on every fact and full history, in one graph.

      Add
    • Catalog and governance

      Still the catalog of record for your tables.

      Keep
    • Lakehouse and warehouse

      Storage and compute stay where they are.

      Keep
    • Source systems

      ERP, CRM, billing and documents stay put.

      Keep
    A typical data and AI stack with Fluree added as one tier. Agents, apps, search and BI sit above it; the catalog, the lakehouse or warehouse and the source systems sit below it and keep their jobs.
    Components

    One context layer, one graph.

    Fluree AI is the product: Resolve, Extract and Model, on the FlureeDB engine. Your agents, apps and people work on top.

    Your workAgents, apps, people
    AgentsAppsPeople
    Fluree AIThe context layer
    ResolveRecords into golden recordsExtractDocuments into graph factsModelContext layer managementFlureeDBThe engine: knowledge graph database
    Your dataRead in place
    Documents & mediaLakehouseDatabases & filesBusiness apps
    • The product

      Fluree AI

      The context layer for enterprise AI. Ask questions, build dashboards and run governed agents, with no infrastructure to run. Its capabilities, Resolve, Extract and Model, all run on the FlureeDB engine. Fluree Memory keeps what agents learn and decide in the same graph, under the same policy as your data, and also ships as a free, source-available tool for coding agents that runs in your repo.

      Product page
    • Entity resolution

      Resolve

      Tables are mapped into the graph with a W3C R2RML mapping. Records are classified against your ontology and resolved into golden records with lineage to every source. Ambiguous matches go to a steward for review.

      Capability page
    • Document extraction

      Extract

      Documents become graph facts, each stored with the excerpt that proves it. Audio and video on the Fluree Enterprise tier.

      Capability page
    • Context layer management

      Model

      Taxonomies and ontologies, built by your experts, reviewed and versioned like a release, then published to the graph. Tables and documents are classified against the same concepts.

      Capability page
    • The engine

      FlureeDB

      The verifiable knowledge graph database under every capability. It holds the graph, enforces policy on every fact, keeps every version and runs search inside the engine. Developers can also run it on their own.

      Engine page
    Data movement

    What moves, and what stays put.

    Tables are read where they live. Documents become facts in Fluree. Nothing is written back.

    • Iceberg tables

      In place, or a verified twin

      Read in place through a W3C R2RML mapping. Optionally materialized into a verified native twin, kept fresh by polling against the last snapshot id, every 30 seconds by default.

    • Delta Lake

      In place

      S3, ADLS, Fabric OneLake, and Databricks via Unity Catalog

      Read in place only.

    • SQL databases

      Live, never copied

      Postgres, MySQL, SQL Server, Oracle, Snowflake, BigQuery and Redshift through a Trino or Starburst endpoint, or Postgres, MySQL and SQLite through the fluree-sql-bridge sidecar

      Read live on every query. Never copied.

    • Documents

      Facts stored, file referenced

      PDF, Word, PowerPoint, Markdown, HTML and images

      Parsed into graph facts that are stored in Fluree. The original file is referenced by name, path and SHA-256 hash, not stored. Re-run ingest to refresh.

    • SaaS apps

      Hosted connectors

      Salesforce, HubSpot, Google Drive and others

      Connected through the connectors in hosted Fluree AI.

    • What Fluree doesn’t do

      There is no change data capture.

      Fluree never writes back to a source.

    Graph sources in the docsDocument pipeline in the docs
    Interfaces

    Standard interfaces, one graph.

    Every interface reads the same governed graph, under the same policy.

    Query languages

    • SPARQL 1.1

      The W3C standard, with zero W3C test-suite failures.

      Docs
    • openCypher

      Property-graph queries and writes over the same graph.

      Docs
    • GraphQL

      An endpoint on every ledger, with the schema derived from the data and its SHACL shapes.

      Docs
    • JSON-LD query

      Fluree’s native query language, written in JSON.

      Docs

    Ways in

    • HTTP API

      Query, transact and manage ledgers over HTTP with the Fluree server.

      Docs
    • MCP server

      An /mcp endpoint turns a ledger into a tool an agent can call. Each token reaches only the ledgers it may read. Off by default.

      Docs
    • CLI

      One binary to create ledgers, load and query data, and run the server.

      Docs

    FlureeDB also runs as a Rust library inside your own process.

    Identity, policy and audit

    Access is decided inside the engine.

    Policy is data in the graph, checked on every fact before a result reaches a model or a person.

    CallerAgent or personOIDC token, server token or did:key signature
    Inside the engineFlureeDBPolicy, stored as data, is checked on every fact
    ResultModel or userSees only what that identity may see
    Every read takes the same path. The caller’s identity selects the policy, the engine applies it to each fact, and nothing the caller may not see leaves the engine. Every commit is kept, so the same read can be replayed as of any past commit.
    Identity
    Every request carries an identity: an OIDC token checked against your provider’s published keys, a server token your admin issues, or a request signed with a did:key. The Fluree Enterprise tier adds SSO. Docs
    Policy is data
    Role, attribute and relationship rules are stored in the graph beside the facts they govern. The engine checks them on every fact before a result reaches a model or a person, and on every write. Docs
    Immutable commits
    A change is a new commit, and history is never modified. Commits and transactions can be signed with Ed25519 keys, so you can prove who wrote what. Docs
    Time travel
    Query any past state by transaction, timestamp or commit. A past read follows the policy as it stood then, so you can replay exactly what an agent saw. Docs
    Branching
    Branch, merge and rebase a ledger like code, to try a change before it lands. Docs
    Audit
    Policies are data, so a policy change is a commit too. Who could see what, and when, is part of the record. Docs
    Deployment

    Run it hosted, or run it yourself.

    One engine and one graph, three ways to run them.

    Self-hosted

    FlureeDB

    A single binary or Docker image. On-prem and air-gapped with full data sovereignty, or in AWS, Azure or GCP with S3 + DynamoDB for HA production.

    Deployment guide
    Hosted

    Fluree AI

    The full platform, serverless and free to start, with MCP and API access. Nothing to install or run.

    Start free
    Dedicated

    Fluree Enterprise

    The enterprise tier of Fluree AI: single-tenant SaaS with region selection and failover across zones, or a private deployment. SSO, SLAs and support.

    Talk to sales

    Where models run. Document parsing needs no model and no GPU. Embeddings and extraction point at any OpenAI-compatible endpoint, such as Ollama, vLLM or OpenAI, or at managed models on Fluree AI.

    Performance and proof

    Measured, dated and checkable.

    Each number has the date it was measured and a source you can rerun or read.

    • 17.5 ms

      Geometric mean across all 105 SPARQLoscope DBLP queries over 561 million triples. First of seven engines on matched hardware; the next takes 11.5× longer.

      Fluree v4.2.0, measured September 2026: benchmark-db

    • 0.934

      Document parsing on opendataloader-bench, 200 PDFs scored for layout fidelity: the highest of 17 engines. The deterministic engine alone, with no model, scores 0.895. The corpus has no scanned PDFs.

      Measured 7 October 2026: fluree-doc-parse

    • 95%+

      In Fluree’s 2024 study, GraphRAG for GenAI Accuracy, answers over raw relational tables scored 8–15% zero-shot; a governed knowledge graph reached 95%+ once enriched and joined.

      Read the report: GraphRAG for GenAI Accuracy

    Full report, raw timings and reproduction scripts
    License

    Free to run, source on GitHub.

    FlureeDB is source-available under the Business Source License 1.1.

    Free in production

    Run it anywhere

    FlureeDB is free to use in production, including inside your own product. The source is on GitHub.

    Commercial license

    Only for one case

    You need a commercial license only to offer FlureeDB itself as a hosted database service.

    Change date

    Apache 2.0 after four years

    Each release converts to Apache 2.0 on its fourth anniversary.

    Read the licenseFlureeDB on GitHub
    Next steps

    Take it further.

    Bring your stack to a session, try it hosted, or install it yourself.

    With your team

    Book an architecture session

    Walk through your sources, your policy model and your deployment with a Fluree engineer.

    Book a session
    Hosted

    Start free

    Fluree AI is free to start, with no infrastructure to run.

    Start free
    Self-hosted

    Read the docs

    Install FlureeDB and query a knowledge graph in under a minute.

    Getting started

    flur.ee/contact?interest=briefingflur.ee/ailabs.flur.ee/docs/db/getting-started