Grounding clinical AI: every answer needs a source.
Why "the model said so" is never good enough at the bedside, and the five-stage pipeline Sage uses so every answer points back to a document you own.
The problem with fluent answers
Language models are built to produce plausible text. In a hospital, plausible is a hazard. When a doctor asks what the antibiotic protocol is for a post-operative fever, the right answer is the one in your hospital's current protocol, not the most common answer on the internet. Grounding means the model may only answer from documents your institution has approved, and must show which ones it used.
Five stages, not one prompt
Sage treats retrieval as a pipeline. Documents such as protocols, SOPs, formularies and discharge summaries are ingested and chunked with their structure intact. A query runs through hybrid search, combining semantic similarity with exact keyword matching, because drug names and lab codes must match exactly. A reranker orders the candidates. An evidence-verification step checks that each sentence in the draft answer is supported by a retrieved passage. Only then is the answer returned, with the exact passages cited.
When the right answer is 'I don't know'
If verification cannot find support for a claim, the claim is removed. If nothing relevant is retrieved, Sage says so and shows what it searched. A refusal with a search trail is more useful to a clinician than a confident paragraph that cannot be checked.
Your corpus, your boundary
Sage indexes the hospital's own content and content the hospital licenses. It runs on-premise or in a private cloud, so protocols and patient documents stay inside the institution's network.
Key takeaways
- 1Hybrid search matters in medicine: drug names and codes need exact matches, not just similar meaning.
- 2Verify every sentence against retrieved evidence before showing it.
- 3A transparent 'not found' beats an uncited answer.