MVP status¶
Canonical maturity and implementation status. Roadmap: architecture/roadmap.md.
Product focus¶
CAE Copilot — Rocky documentation assistant (portfolio demo).
Corpus: data/cae/ansys_rocky/v26r1/docs/ (9 PDF manuals, private volume).
Maturity stages¶
| Stage | What it does | Status |
|---|---|---|
| 1 — Data Foundation | Rocky v26r1 PDF ingest, page-aware chunks, manifest metadata | Done (MVP corpus) |
| 2 — Trusted Agent Layer | LangGraph, RAG + citations, DEM tools, eval, API/UI | Done |
| 3 — Live CAE integration | Rocky API / job orchestration (license required) | Deferred |
Shipped in prod¶
| Surface | URL |
|---|---|
| UI (Streamlit) | cae-copilot.dmitryivanov.dev |
| API (Swagger) | api.cae-copilot.dmitryivanov.dev/docs |
| Docs | docs.cae-copilot.dmitryivanov.dev |
Rocky v26r1 PDF corpus ingested into Qdrant on healserver. API image bakes
multilingual-e5-base offline — no HuggingFace egress at runtime.
Quality gates¶
mamba activate cae-copilot
pdm run lint && pdm run typecheck && pdm run test
Eval (14 golden cases, rule-based metrics):
pdm run eval
Methodology: tests/eval/run_eval.py invokes the full graph per case in
golden_qa.jsonl. Metrics: intent accuracy (all cases), citation hit rate
(cases with expect_citations), tool accuracy (cases with expect_tool).
Offline baseline (no LLM key, in-memory RAG): 100% / 100% / 100% on the
current golden set. LangSmith/datasets faithfulness checks are planned.
Local compose¶
make dev-stack && make ingest && make smoke
Next¶
- Portfolio demo video per walkthrough
- Expand eval golden set to 15–20 cases (
tests/eval/golden_qa.jsonl) - Optional: quick/deep mode toggle in Streamlit UI (API already supports
mode)