Ask the reference.
Three starter questions are below, or type your own. Answers stay within the retrieved passages and are kept to three sentences.
Hello. Ask me about a condition, a medicine, or a treatment — I answer from the medical reference I was indexed on, and I will tell you when it does not cover something.
POST /get · retrieval first, then a constrained answer · not medical advice
Grounded, not guessing.
A retrieval-augmented generation pipeline. The model never answers from memory alone — the reference is retrieved first, every time.
-
Ingest & chunk
The medical reference PDF is split into overlapping 500-character chunks, so each passage stays small enough to retrieve precisely.
-
Embed
Every chunk becomes a 384-dimension vector with the open all-MiniLM-L6-v2 sentence-transformer.
-
Retrieve
Your question is embedded the same way, and Pinecone returns the three most similar passages by cosine similarity.
-
Answer
An OpenAI model writes a concise answer constrained to those passages — and declines when they do not support one.