Olink Knowledge Graph Platform

A production system for building and querying biomedical knowledge graphs from scientific literature. Extracts protein-disease relationships from papers using LLMs, stores them in graph databases, and exposes a RAG-powered API for intelligent queries.

Architecture Overview

graph TB
    subgraph Sources["Data Sources"]
        PM[PubMed]
        PMC[PMC Full-Text]
        BR[bioRxiv]
        PDF[Local PDFs]
        CSV[CSV/Ontologies]
    end

    subgraph Ingestion["Ingestion Pipeline (bioingest)"]
        FT[Fetch & Extract Text]
        CH[Token Chunker]
        LLM[LLM Entity Extraction]
        SC[Schema Validation]
        ER[Entity Resolution]
    end

    subgraph Storage["Storage Layer"]
        NP[Neptune / Neo4j<br/>Graph Database]
        AU[Aurora pgvector<br/>Embeddings]
    end

    subgraph Serving["Query Serving (graphrag-api)"]
        API[REST API<br/>FastAPI + Granian]
        AG[Query Agents<br/>Two-Phase / Dynamic]
        KT[KG Tools<br/>Cypher + Vector Search]
    end

    subgraph Clients["Clients"]
        FE[React Frontend]
        PL[00panel Dashboard]
        CL[curl / SDK]
    end

    Sources --> FT
    FT --> CH --> LLM --> SC --> ER
    ER --> NP
    CH --> AU

    NP --> KT
    AU --> KT
    KT --> AG --> API

    API --> FE
    API --> PL
    API --> CL

    style Sources fill:#e8f4fd,stroke:#2196F3
    style Ingestion fill:#fff3e0,stroke:#FF9800
    style Storage fill:#e8f5e9,stroke:#4CAF50
    style Serving fill:#f3e5f5,stroke:#9C27B0
    style Clients fill:#fce4ec,stroke:#E91E63

How It Works

Step	What happens	Where
1. Fetch	Papers downloaded from PubMed/PMC/bioRxiv or PDFs loaded	bioingest
2. Chunk	Text split into 512-token segments with overlap	bioingest
3. Extract	LLM (Bedrock) identifies proteins, diseases, relationships	bioingest
4. Validate	Schema enforces entity types and filters garbage	bioingest
5. Resolve	Duplicates merged via UniProt ID, MONDO ID, fuzzy matching	bioingest
6. Store	Nodes/edges → Neptune, embeddings → Aurora pgvector	bioingest
7. Query	User asks question → agent searches graph + vectors → LLM synthesizes answer	graphrag-api

Repositories

Repo	Role	Key Entry Points
graphrag-api	Query serving, REST API, agents	api/app.py, src/agents/
bioingest	Data fetch + KG creation	uv run bioingest ingest
graphrag_eval	Benchmarking + evaluation	uv run python -m benchmark
graphrag_react	React frontend	npm run dev

Quick Start

aws sso login
cd graphrag-api && uv sync
uv run python ingest_main.py --source pdf --pdf-files paper.pdf --service bedrock
uv run granian --interface asgi api.app:app --host 0.0.0.0 --port 8000

See Getting Started for the full setup guide.

Quick Navigation

Section	Description
Getting Started	Installation, configuration, first KG build
Build Your First KG	15-minute hands-on walkthrough
CLI Reference	All commands and flags
Architecture	System design, data model
Schema Expansion	PTMs, variants, data model discussion
Ingestion Pipeline	Sources, extraction, consolidation
Query System	Agents, modes, search depth
API Reference	REST endpoints, SSE streaming
Neptune	Production ingestion runbook
Operations	Maintenance, incremental updates

Production Stats

Backend	Nodes	Relationships	Embeddings
Neptune (beta)	70K	263K	—
Aurora pgvector	—	—	216K

Tech Stack

• API: Python 3.12 / FastAPI / Granian (Rust ASGI)
• Graph: Neptune (serverless) / Neo4j
• Vectors: Aurora pgvector (768-dim, cosine)
• LLM: AWS Bedrock (Llama 3.3, Nova Pro, Claude)
• Frontend: React + Sigma.js
• Infra: AWS CDK / ECS Fargate / EventBridge
• CI: GitHub Actions (two-layer Docker: base cached, app fast)