The query "healthy high-protein chicken bowl" has no exact match in any recipe title, which is why keyword search fails at it. Hybrid retrieval with reranking handles the intent, and the diet, time and difficulty filters apply on top — so what comes back is both relevant and cookable tonight.
A hosted MCP server doing semantic search over the Food.com recipe collection, at https://recipes.aidatanorge.no/mcp. Retrieval is two-stage: dense embeddings plus sparse BM25, fused with reciprocal rank fusion and then reranked by a cross-encoder.
- Natural-language search across the whole recipe collection, in English or Norwegian, without keyword syntax
- Filters that match how people decide: diet (vegetarian, vegan, gluten-free, dairy-free, low-carb, keto, paleo), maximum total minutes, and difficulty
- Full records back, not just titles — ingredients, instructions, servings, total time, nutrition, rating and rating count
- Ranking scores returned with each hit, so the ordering is inspectable rather than opaque
- Nutritional detail per serving: calories, fat, protein, carbohydrates, sodium, fibre and sugar
- Five results per query by default, and more on request up to a fixed ceiling
Nothing to install — a streamable-http entry pointing at the hosted endpoint, or a single `claude mcp add` line. No account and no key. A browser demo runs at the same domain if you want to try the search before wiring it in.
One command — claude mcp add --transport http food-recipe https://recipes.aidatanorge.no/mcp