A temporal knowledge graph you can run anywhere
Predicato is an open-source framework for building knowledge graphs that evolve over time. Extract entities and relationships from documents using local ML models, store them with full temporal context, and query across your entire knowledge base — all without sending a single byte to an external API.
Go library with a Python client. Embedded databases. Local ML via Rust FFI. No API keys, no vendor lock-in, no recurring costs.
View on GitHub Learn MoreWhy Predicato
Most knowledge systems require cloud APIs for every extraction and query. Predicato runs entirely on your hardware — no API keys, no per-call costs, no data leaving your network.
Fully local ML stack
Embeddings (Qwen3-0.6B), text generation (SmolLM2), NER (GLiNER), and reranking all run locally via Rust FFI. Models auto-download from HuggingFace and cache locally. Zero external API calls required.
Bi-temporal knowledge
Every fact tracks two time dimensions: when it was recorded and when it was actually true. Query "what did we know as of last Tuesday" or "show me facts that were later corrected." Essential for compliance and audit trails.
No recurring costs
No per-query fees. No per-extraction charges. No monthly API bills that scale with usage. Extract once, query forever. Your knowledge, your infrastructure, your budget.
Two-layer architecture
Predicato separates expensive extraction from flexible modeling — extract facts once, build multiple graph views without re-processing.
Layer 1: Fact Store
Documents, conversations, and events are processed through entity extraction (GLiNER + NLP) to produce entities, relationships, embeddings, contextual triples, and conditional rules. These persist durably in PostgreSQL with vector embeddings — extraction is expensive, so you only do it once.
Backends: PostgreSQL + VectorChord (production), DuckDB (embedded)
Layer 2: Knowledge Graph
Facts from the store are resolved (entity deduplication, relationship merging), enriched with community detection, and assembled into a queryable graph. Regenerate the graph with different parameters without re-extracting — experiment freely.
Backends: CozoDB, DuckDB+DuckPGQ, Ladybug (embedded); Neo4j, Memgraph (external)
Example: Payer contract knowledge graph
Extracted rules
THEN prior authorization required
UNLESS emergency admission
Source: contract.pdf, page 14 · Valid: 2026-01-01 → 2026-12-31
THEN HbA1c testing covered quarterly
Source: contract.pdf, page 22 · Valid: 2026-01-01 → 2026-12-31
THEN reimbursement = $130.08
UNLESS bundled with same-day procedure
Source: fee_schedule_2026.pdf, page 3
Temporal facts
What it does
Extract structured knowledge from unstructured sources. Search across it intelligently. Keep it current as the world changes.
Document extraction
Feed in PDFs, text, HTML, or any document. Predicato extracts entities, relationships, contextual triples (with conditions, temporality, certainty, and scope), and conditional rules (IF-THEN-UNLESS patterns) — all with source attribution.
Hybrid search
Combine semantic similarity (cosine on embeddings), BM25 keyword matching, and graph traversal in a single query. Five reranking strategies: RRF, MMR, cross-encoder, node distance, and episode mentions.
Entity resolution
Automatically identify and merge duplicate entities across sources. "Dr. Smith," "John Smith MD," and "J. Smith, Internal Medicine" resolve to the same node in the graph.
Incremental updates
Add new documents to an existing graph without reprocessing everything. Predicato resolves new entities against existing ones and integrates new relationships into the graph incrementally.
Community detection
Discover clusters of related entities automatically. Useful for identifying topic areas, organizational structures, or conceptual groupings within your knowledge base.
Conditional rules
Predicato extracts IF-THEN-UNLESS patterns from documents: "If CPT 70553 AND outpatient setting, THEN prior auth required, UNLESS emergency." These rules are queryable and auditable.
Pluggable everything
Start embedded, scale to external services. Swap any component without changing your application code.
Graph databases
Embedded: CozoDB, DuckDB+DuckPGQ, Ladybug
External: Neo4j, Memgraph
Embeddings
Local: go-candle (Qwen3-0.6B)
Cloud: OpenAI, Gemini
Text generation
Local: go-candle (SmolLM2-360M)
Cloud: OpenAI, any compatible API
Entity extraction
Local: GLiNER (ONNX)
Cloud: GLiNER2 API, LLM-based
Use cases
AI agent memory
Give your AI agents persistent, queryable memory that tracks what they've learned over time. Multi-turn conversations become knowledge that persists across sessions.
RAG systems
Go beyond vector search. Predicato's hybrid search (semantic + keyword + graph traversal) with 5 reranking strategies produces more relevant retrieval than embeddings alone.
Compliance & audit
Bi-temporal tracking means you always know what you knew and when. Source attribution traces every fact back to its origin document. Essential for regulated industries.
Offline deployments
Air-gapped environments, on-premises installations, or simply reducing cloud spend. Predicato's embedded stack means no internet required after initial model download.
Production ready
Built for real workloads, not just demos.
Circuit breakers & fallbacks
If a cloud provider is configured and fails, Predicato falls back to local models automatically. Token usage tracking and cost calculation built in.
Multi-tenancy
GroupID-based isolation lets multiple applications or organizations share a single Predicato instance without data leakage.
WAL recovery & telemetry
Write-ahead logging for embedded databases. Telemetry with DB persistence for operational monitoring. MCP server support for tool integration.
