About this project¶
AI Engineering Portfolio — RAG architecture demo with licensed CAE manuals as private test data.
Problem and scope¶
See Problem statement for the Rocky engineer pain points, five engineering problem areas, and in/out-of-scope table. This page focuses on what shipped in the demo build.
Solution¶
A Retrieval-Augmented Generation (RAG) system demonstrating chunking, embeddings, vector retrieval, LangGraph orchestration, and grounded answer synthesis with numbered citations and confidence analytics. The focus is the pipeline, not the vendor.
Tech stack¶
| Layer | Technology |
|---|---|
| API | FastAPI, Python 3.11 |
| UI | Streamlit (chat + legal; HTTP client only) |
| Orchestration | LangGraph |
| LLM | OpenRouter / OpenAI-compatible API |
| Embeddings | sentence-transformers (multilingual-e5-base) |
| Vector DB | Qdrant |
| Ingest | PyMuPDF, parent/child chunking, manifest metadata |
| Deploy | Docker, Kubernetes, GitLab CI |
Features¶
- Semantic search over Rocky v26r1 PDF manuals
- Source attribution with page numbers and citation cards
- Retrieval-augmented generation with confidence scoring
- PDF ingestion pipeline via manifest
- Intent routing (knowledge / setup / scripting / calculation / unknown)
- Three deterministic DEM helper tools (particle count, unit convert, timestep heuristic)
- SSE streaming (
POST /ask/stream) with pipeline stage indicators - Follow-up question chips and starter prompts (
GET /suggest-questions) - Thumbs up/down feedback (
POST /feedback) - Scripting intent: copy/download Python draft blocks
- Regenerate answer; rate limiting; demo kill switch (
CAEC_DEMO_ENABLED)
Related docs¶
- RAG pipeline — layer diagram and technologies
- Knowledge base — corpus, ingest, live stats API
- Disclaimer — legal notice (also via Streamlit Legal panel)