Corpus

Evidence quality starts before the model.

AskStein's orthopedic corpus is acquired, classified, filtered, and converted into traceable evidence units before it is ever used to compose an answer.

17.7K

Papers acquired & classified

16.4K

Retained at ≥0.7 relevance

~12K

Verified Q&A pairs · target

At a glance

Acquired: 17,700 orthopedic papers acquired and classified

Retained: 16,400 papers retained at a relevance score of 0.7 or higher

Training set: Approximately 12,000 verified question-and-answer pairs targeted

Specialty: Orthopedic clinical decision support (the current beachhead)

Embeddings: OpenAI text-embedding-3-large

Index: FAISS, built for fast similarity search

Traceability: Answers link back to the supporting paper span and PMID

How the corpus is built

Ingestion & curation

Papers are acquired from approved channels, classified for orthopedic relevance, deduplicated, and split into evidence passages.

Embedding

Each passage is converted into a vector using OpenAI's text-embedding-3-large.

Indexing

Those vectors are stored in a FAISS index built for fast similarity search.

Acquire → process → fine-tune → publish

The pipeline is designed as a closed evidence loop: acquire literature, process it into traceable units, use verified question-and-answer pairs to improve the response layer, and publish only answers whose supporting evidence can be resolved to a source.

Why the corpus is the moat

AskStein's central bet — confirmed empirically in its own testing — is that retrieval quality, not model size or fine-tuning, is the dominant driver of answer accuracy. Improving the corpus and the retrieval pipeline moves the needle more than swapping in a bigger model. Anyone can call an LLM; the defensible asset is the curated body of evidence and the standards that keep it clean.

See how the corpus is actually used at query time in the documentation →

Where it scales next

The orthopedic corpus is the first vertical. The next phase is designed to ingest PubMed at scale — more than 40 million abstracts, with full text included where it is openly licensed — and extend the same evidence and abstention discipline across specialties.

Every answer should lead back to the exact evidence that supports it.