RAG pipeline¶
End-to-end flow from PDF manuals to cited answers in the portfolio demo.
flowchart TD
PDF[PDF Documents] --> Chunk[Chunking]
Chunk --> Embed[Embeddings]
Embed --> VDB[Vector Database]
VDB --> Ret[Retriever]
Ret --> LG[LangGraph Orchestration]
LG --> LLM[LLM]
LLM --> Out[Answer + Citations]Layers¶
| Layer | Technology | Purpose |
|---|---|---|
| PDF Documents | PyMuPDF | Ingest CAE manuals (page-aware text extraction) |
| Chunking | Parent/child splitter | Searchable child chunks with metadata |
| Embeddings | sentence-transformers (multilingual-e5-base) | Semantic vectors for retrieval |
| Vector Database | Qdrant | Persistent index with payload filters |
| Retriever | Cosine similarity + parent expansion | Top-k retrieval with provenance |
| Orchestration | LangGraph | Intent routing → RAG / tools → cited report |
| LLM | OpenRouter / OpenAI-compatible | Grounded synthesis with numbered citations |
| Answer + Citations | FastAPI + Streamlit | Traceable responses, SSE streaming, analytics |
The same layer list is exposed at runtime:
curl https://api.cae-copilot.dmitryivanov.dev/architecture