Flexible search infrastructure
The retrieval pipeline can combine a local corpus and free bibliographic databases without requiring a paid search API. This does not mean the entire platform runs without external services. Inference, embeddings, or optional features may use configured providers.
Retrieval sources
| Layer | Examples | Role |
|---|---|---|
| Local | Institutional corpus, documents, and indexed guidelines | First option when relevant content is available. |
| Bibliographic | PubMed, Europe PMC, OpenAlex, and Lens | Biomedical research using redacted queries. |
| Governed web | SearXNG or enabled providers | Fallback conditioned on policy, query type, and budget. |
Separate search from inference
A free search can feed a paid remote model. A local model can query an external source. Cost and data sharing therefore need to be assessed at each stage rather than through one infrastructure label.
Documents and RAG
Documents can go through extraction, chunking, BM25 indexing, and vector indexing. OCR, Qdrant, embedding models, and re-ranking depend on environment configuration.
Deployment review
Before a pilot, confirm active sources, local and remote components, data received by each provider, costs and limits, and behavior when a provider fails.