Knowledge base¶
The demo indexes ANSYS Rocky v26r1 manuals from a private local volume. PDFs are never committed to git or baked into public Docker images.
Corpus (operator machine only)¶
Nine Rocky v26r1 PDF manuals are declared in
data/cae/ansys_rocky/v26r1/docs/manifest.yaml (see manifest.example.yaml in git).
After a full local ingest, expect on the order of ~3,400 pages and ~15,000 chunks
(embedding model: multilingual-e5-base).
Ingest¶
make dev-stack
make ingest # requires manifest.yaml + PDFs on disk
See quickstart.md and deployment for paths and Helm ingest job (prod: Rocky PDFs from Nextcloud data/cae volume).
Live statistics¶
curl http://localhost:8000/kb/stats
# or production:
curl https://api.cae-copilot.dmitryivanov.dev/kb/stats
Example response fields:
| Field | Meaning |
|---|---|
documents |
Number of indexed PDFs |
pages |
Total PDF pages with chunks |
chunks |
Child chunks in Qdrant |
embedding_model |
Active embedding model id |
collection |
Qdrant collection name |
cae_product |
Adapter id (e.g. ansys_rocky) |
version |
Corpus version tag (e.g. v26r1) |
Privacy¶
Manuals are loaded from Nextcloud-synced or operator-local paths only. The public Streamlit UI is chat + legal; corpus details live in this documentation.