Restaurant Menu Retrieval Contracts: Python Metadata Filters at Scale

Restaurant menu retrieval contracts emphasize the importance of defining a retrieval contract first, utilizing durable menu text within a vector index, and incorporating live retrieval only for facts whose freshness is crucial. Measuring re-index work, filtered recall, and citation coverage is essential before selecting a backend.

The assistant should provide more than a generated sentence; each hit must include an item identifier (or source URL), restaurant, location, menu version, dietary labels, and an updated timestamp. The answer layer can then display information such as "vegan, Midtown, dinner menu" with evidence rather than requiring the model to remember where a claim originated.

Metadata is treated as a product interface, with filter parameters including restaurant_id, location_id, meal_period, allergens, availability, and menu_version. When querying for a "nut-free lunch near SoHo," semantically similar but inappropriate dinner items should be excluded before presenting the model with the results. This approach proves more cost-effective than retrieving multiple plausible but unusable chunks and spending tokens explaining the exclusions.

A vital operational rule is to set a timeout, retry budget, and result limit for each source. Slow web pages must not hinder checkout. When a menu changes, re-index the modified records explicitly; when an item is deleted, remove its vector rather than leaving a stale answer available. This requires tracking an ingestion event, comparing the menu version with the stored version, and integrating the delete path into the same runbook as upsert.

Otherwise, a "gluten-free" response could be constructed from yesterday's item despite the current menu being accurate in other respects, and a later model prompt might conceal the error behind fluent prose. Metadata filters should remain simple. Python code facilitates the durable path of retrieval contracts, handling vector service collection creation, upsert operations, and vector queries.

The example code keeps the contract visible without assuming a universal schema for embedding models. The vector service is intentionally excluded from the example; vectors are produced by an evaluation harness and fed into the index.