Preciso turns unstructured documents into structured, queryable knowledge graphs — without a fixed extraction pipeline baked into the server. Instead, an agent (with its own file tools and an LLM) reads a skill — a plain-markdown spec describing what entities and relationships matter for a given domain — and writes structured extractions to disk. The MCP server never parses the domain logic itself; it only ingests the resulting extraction JSON, reconciles it against the graph, and serves it back for querying.

This split keeps the core engine small, auditable, and domain-agnostic, while the extraction logic stays swappable, versionable, and community-owned.

Benchmark: 95.4% accuracy with zero hallucinations across 23 test cases on a Walmart 10-K extraction benchmark.

                 ┌─────────────────────┐
   documents ──► │   skill (markdown)   │  ← domain-specific extraction spec
                 │   read by an agent    │     (financial, general, research…)
                 └──────────┬───────────┘
                             │ writes extraction JSON
                             ▼
                 ┌─────────────────────┐
                 │   preciso-graphrag    │  ← MCP server: ingest, reconcile,
                 │   (core engine)        │     store, query the graph
                 └──────────┬───────────┘
                             │ MCP tools
                             ▼
                 ┌─────────────────────┐
                 │       Preciso          │  ← web frontend + visualizer
                 │   (marketing + UI)      │     (d3-force graph, streaming LLM)
                 └─────────────────────┘

• The graph creation process is the tool. Extraction isn't a hidden pipeline — it's a markdown skill the agent can read, reason about, and follow, the same way a human analyst would follow a style guide.
• Provider-agnostic. Bring your own LLM, embedding model, and coding agent. Preciso doesn't lock you into a single stack.
• File-based handoff. Agents write extractions to disk; the server ingests and reconciles. This keeps agent reasoning and server logic decoupled and independently testable.
• Skills are content, not code. New domains ship as .md files, not server changes — anyone can contribute a skill without touching the core engine.

git clone https://github.com/Preciso-GR/preciso-graphrag.git
cd preciso-graphrag
# see the repo README for MCP server setup and skill configuration

Point your .mcp.json at the server, drop a skill from skills_for_preciso (or write your own) into skills/, and start extracting.

Skill contributions are the easiest way to extend Preciso to a new domain — see skills_for_preciso for the format and examples (financial-graph-extraction, general-graph-extraction, research-paper-graph-extraction). Core engine contributions and issues go to preciso-graphrag.