Poultry Health Advisor · Compare
The same question, with and without the corpus.
Both answers come from the same model (claude-sonnet-4-6). The grounded side feeds the retrieved passages to it, so it costs a little more per answer, still cents, and refuses off-scope questions for nothing. What you buy for that is proof: every claim cited to a paper and page you can open and check. A general LLM with no corpus answers from memory, cannot cite, and does not know when a question is outside its remit.
I · Ask both
II · Side by side
No corpus
General-purpose LLM
Pick a question above to see the ungrounded answer.
Corpus-grounded
poultry-poc
Pick a question above to see the grounded answer.
Retrieval match strength is the similarity of the retrieved passages, a heuristic operating point, not a calibrated confidence in the answer: a grounded deferral can show a strong match when relevant passages were retrieved that do not cover the specific question. Calibrating these thresholds against a labeled set, with a reported false-accept and false-refuse rate, is part of the funded work.
III · What the comparison shows
Cost. Same model, but the grounded side feeds the retrieved passages to it, so a cited answer runs about two cents against roughly one cent for the bare model, a cent or so more, still cents. On an off-scope question the balance flips: poultry-poc refuses before the model is ever called, for nothing, while a general LLM always pays to answer something outside its remit. Which guard does the refusing is named under each answer, and counted per mechanism on Methods.
The difference that matters is not price. Every poultry-poc claim is cited to a specific paper and page; it refuses when the corpus does not cover the question instead of guessing; and every citation is verifiable against the source panel. The general LLM can do none of these, not because it is a worse model, but because it has no corpus to ground, cite, or gate against.