Embedded intelligence that never leaves your building.
Datagraph is a fully local, agentic RAG platform for supply chain operations. It ingests your documents and structured data, connects them in a knowledge graph, and answers questions with grounded, traceable, source-cited intelligence — running entirely on your own infrastructure.
Every model, every embedding, every query runs on-premise. No data ever crosses your network boundary.
Four retrieval paths — keyword, vector, graph, and reasoning — fused and reranked so answers come from your data, not model memory.
A live reasoning trace shows every step the system takes on your behalf. A glass box, not a black box.
Continuous evaluation, golden datasets, SQL governance and corpus health monitoring — quality you can measure, not assume.
Quad-path hybrid retrieval
Vector search alone misses exact terminology, multi-hop relationships, and structure. Datagraph runs four complementary retrieval paths in parallel, fuses their results, and reranks with a cross-encoder — one pipeline that covers the failure modes of each individual method.
Keyword — Hybrid
Catches exact terminology: “OTIF”, incoterms, part numbers, clause references.
Vector — pgvector
Semantic similarity via bge-m3 embeddings; finds meaning beyond wording.
Graph — Retrieval
Traverses supplier → event → plant relationships that span multiple documents.
Reasoning — PageIndex
Navigates long contracts and reports by structure, the way an expert reads.
Metadata pre-filtering scopes every path before it runs — the right candidates, ranked by relevance, delivered with citations.
Nine platform innovations
Beyond the retrieval pipeline, Datagraph introduces platform-level capabilities designed for one purpose: enterprise supply chain intelligence you can trust, verify, and govern.
Multi-dimension metadata model
Every document and chunk is classified along Industry-specific schema-driven dimensions — knowledge layer, topic, scope, lifecycle, strategic intent. Adding a dimension is a JSON edit, not a code change.
Graph-enhanced knowledge
Entities and relationships are extracted at ingestion and stored as a living knowledge graph — enabling multi-hop answers no single document contains.
NL2SQL structured data agent
Plain-language questions become validated SQL against your live supply chain schema — 60+ tables of masterdata, orders and BOMs — with governance, benchmarks and full traceability.
Live reasoning trace
Every pipeline step — query rewrite, retrieval path, graph traversal, SQL generation, sub-agent spawn — streams to the user in real time. Trust through visibility.
Three-layer memory
Conversation history, session scope and long-term user context persist across queries — so a follow-up question never needs the full context restated.
Sub-agents with isolated context
Specialist agents — document analyst, SQL analyst, market researcher — run in isolated context windows and report back, keeping the main thread clean and focused.
Evaluation & corpus health
Golden datasets, four-layer evaluation, regression detection and corpus health KPIs — is the corpus complete, fresh, trusted, used and reliable? Now you know.
External signal & market agents
A News Agent and Market Analyst monitor the outside world across 30+ signal categories, feeding a ranked signal feed and structured intelligence reports into the graph.
Workspaces & domain harnesses
Project workspaces with Deep Mode: choreographed multi-phase workflows for SQL analysis and supply chain contract review, producing board-ready reports.
Where does your intelligence live?
Most AI platforms send your contracts, supplier data and strategy documents to someone else’s cloud. Datagraph takes the opposite position: the models come to your data, not the other way around.
Local LLMs, embeddings and reranking on your hardware, one PostgreSQL instance for vectors, graph, tables and auth. Even voice and document parsing have a fully local path. Sovereignty isn’t a feature tier — it’s the architecture.
Technology foundation
Datagraph is built on best-in-class open technology — deliberately chosen so the entire stack can run locally, be audited, and never lock you into a single vendor.
| Technology | What it does | Why it matters | Example |
|---|---|---|---|
| Local LLMllm runtime | Runs open large language models entirely on local hardware — classification, generation, extraction. | Full data sovereignty; zero external API calls at runtime. | ollama.com |
| bge-m3embeddings | Multilingual embedding model converting text into vectors for semantic search. | State-of-the-art open embeddings; 100+ languages, long-context. | BAAI · Hugging Face |
| BGE Rerankercross-encoder | Re-scores the fused candidate set by reading query and passage together. | Sharply improves precision of the final context window. | BAAI · Hugging Face |
| GraphRAGgraphrag | Extracts entities and relationships from chunks and retrieves over the resulting knowledge graph. | Multi-hop answers across documents; runs as a local sidecar. | HKU · GitHub |
| Indexreasoning retrieval | Builds a hierarchical tree of long documents and navigates it with LLM reasoning — vectorless, chunkless. | Expert-grade precision on contracts, filings and 50+ page reports. | VectifyAI · GitHub |
| Doclingdocument parsing | Layout-aware parsing of PDF, DOCX, XLSX, PPTX, HTML and scans (OCR) into structured text. | One open parser for every enterprise format, including tables. | IBM Research · GitHub |
| Supabasepostgres platform | PostgreSQL with pgvector, row-level security, auth, storage and realtime — the platform’s single database. | Vectors, graph tables, business data and auth in one auditable instance. | supabase.com |
| pgvectorvector search | Native vector similarity search inside PostgreSQL. | No separate vector database to operate, sync or secure. | GitHub |
| RRFrank fusion | Reciprocal Rank Fusion merges ranked results from all four retrieval paths. | Robust, parameter-light fusion proven in IR research. | Cormack et al., 2009 |
| FastAPI + SSEbackend | Python backend streaming reasoning steps and answers as Server-Sent Events. | Real-time transparency with a lightweight, standard protocol. | fastapi.tiangolo.com |
| React + shadcn/uifrontend | TypeScript frontend with Vite, Tailwind and shadcn/ui components; Recharts for analytics, Leaflet for maps. | Fast, accessible, consistent UI across chat, dashboards and maps. | ui.shadcn.com |
| Whispervoice | Speech-to-text and text-to-speech layer wrapped around the unchanged chat pipeline. | Hands-free operation; local path via Whisper.cpp + Kokoro TTS. | OpenAI · GitHub |
| APSchedulerautomations | Runtime scheduler behind the unified Automations control center — evals, agents, reports. | Every automation visible, schedulable and auditable in one place. | apscheduler.io |
// entire runtime stack available under open licences — cloud connectors used only during development, fully local at deployment