🚧 Under construction

Graph-augmented semantic search for academic literature

Current frontend with mock data: semantic results + graph view

• 📄 Fetch papers from ArXiv API
• 🧠 Queue papers for embedding (deferred, async)
• 🧮 Track paper ingestion state in SQLite index
• 📦 Index embeddings into Qdrant
• 🔍 Search Qdrant with fastembed support
• 🧩 Merge vector + (planned) graph hits
• ✅ Type-safe, testable, modular pipeline
• ⚙️ Health checks, typed interfaces, and task runner setup

• Python 3.12+
• Node.js 20+
• Docker + Docker Compose

make up-full # Run full stack with CPU make up-full USE_GPU=1 # Run with GPU (requires NVIDIA runtime)

Services:

• API → http://localhost:8000/docs
• Qdrant UI → http://localhost:6333/dashboard
• Redis → localhost:6379 (use redis-cli)

poe server # Run the FastAPI backend poe pipeline # Run the end-to-end pipeline (logs steps) poe test # Run the tests

flowchart TD subgraph API A1[GET /search] --> P["run_pipeline()"] end subgraph Pipeline P --> F[Discover papers from ArXiv] F --> I[Update PaperIndex] I --> Q[Enqueue papers if not embedded] Q --> S[Semantic Search] S --> G[Get related from GraphStore] G --> M[Merge vector + graph results] M --> R[Return SearchResults] end subgraph Vector Store V1["Qdrant (hosted/local)"] end subgraph Embedding Worker W1["Reads Redis queue"] W1 --> E[Embed papers] E --> V[Upsert to Qdrant] V --> U[Update PaperIndex status] end subgraph Graph Store G1["(Planned) Neo4j / in-memory graph"] end S -->|vector hits| V1 G -->|edges| G1 G1 -->|related| G 

Stack highlights:

• Backend: FastAPI + Pydantic + Poetry (with poethepoet task runner)
• Queue: Redis (Upstash or local)
• Vector DB: Qdrant
• Frontend: React + Vite + Tailwind
• Infra: Docker Compose + Fly.io

Thank you to arXiv for use of its open access interoperability.

MIT