ask_about_restaurant
ChatGPTAsk a specific question about a restaurant based on analysis of real reviews and menu data. Common questions: what to order, group suitability, dietary options, vibe/atmosphere, value assessment. Requires a restaurant_id from find_restaurant or search_restaurants. Ask one question per call. Accounts connected through an assistant have a small free number of answers per day; if the result carries limit_reached, relay its message to the user as written and do not retry.
explore_area
ChatGPTGet aggregate dining statistics for a neighborhood, city, or region: cuisine breakdown, grade distribution, price range, top neighborhoods, and highlighted restaurants. Use this for area-level context ("what is the dining scene like in Shoreditch?"), NOT for finding a specific restaurant or getting a personal recommendation. For "find me a quiet Italian near Shoreditch", use the recommend tool instead. Coverage is reported honestly: coverage_level 'full' (fully analyzed rows present), 'basic' (only quick-read rows: review-analysis bands, no letter grades yet), or 'none'; analyzed_count and preliminary_count split total_restaurants, and highlights marked coverage_level 'basic' carry a preliminary_band instead of a grade; the band is the floor of the range, so 'B-range' means at least B minus. not_found means neither analyzed nor preliminary rows exist.
find_restaurant
ChatGPTLook up restaurants by name when the diner, a booking app or a guide has named one: returns matching places with Seemor IDs, grade or preliminary band, a one-line read and seemor_url. Pass city whenever the conversation shows it, in English or the local language; a neighborhood also works. Partial names match; chains return one row per branch. Not for choosing where to eat: to pick a place for a meal, call recommend with the ask instead of inventing candidates and checking them here. Each result reports coverage_level: 'full' (letter grade), 'basic' (quick read: grade null, preliminary_band which is the floor of the range, tldr, caveat) or 'none' (no analysis yet). A 0-result answer means Seemor has no match for that spelling; try the name without a leading article (Le, La, Cafe, Chez) before telling the diner it is not covered. Rows carry google_maps_url and website when Seemor holds them (null otherwise), plus seemor_url. Seemor holds no booking, availability or opening-hours data: use a booking app for those and join on restaurant name and address.
lookup_restaurant
ChatGPTLook up a specific restaurant by its Seemor ID. Returns grade, summary, cuisine, neighborhood, and other details. Use the fields parameter to request richer data (standard or premium; fully analyzed restaurants only, except the quick-read dimension_assessments below). coverage_level 'full' rows carry a letter grade; 'basic' rows are Seemor quick reads: review-analysis bands (grade null, preliminary_band such as 'B-range') with a one-line tldr, graded from review analysis rather than star ratings; the band is the floor of the range, so 'B-range' means at least B minus. With fields 'premium' a quick-read row also carries dimension_assessments for the dimensions the quick read measured (food quality, consistency, noise, service attentiveness, large parties, local vs tourist, value), labelled on the same scale as full analysis but read from fewer reviews, so present them as preliminary. 'none' rows have no analysis yet. Rows carry google_maps_url and website when Seemor holds them (null otherwise), plus seemor_url. Seemor holds no booking, availability or opening-hours data: use a booking app for those and join on restaurant name and address. Use search_restaurants or find_restaurant first to get restaurant IDs. Use this for a single place the user asks about, not for every result of recommend.
recommend
ChatGPTThe one call for "where should we eat": a ranked, final answer for one meal or occasion, each result graded with the reasons it fits and anything it misses. Call it once per meal slot with the diner's ask in the diner's own words: the place or anchor, the time of day, who is eating (party size, kids, a date, a group) and any stated constraints or dislikes. Prefer it over generating your own candidates and checking them by name; it already ranks what Seemor holds for the ask. A dish ask ("good beef bourguignon near the Marais") is a recommend ask. Results are ranked and final: do not re-call with reworded variations or look up each result. Each result already carries a grade (or preliminary_band, the floor of the range: 'B-range' means at least B minus), match_reasons, caveats (a caveat beginning "Misses" names something the room does not deliver, never a good thing about it; a strong_match row never carries one), honored_constraints (plain-language labels for the stated filters this specific row confirms, e.g. "Indian", "Takeaway", "In Magnolia"), and requirements (per stated requirement, one of confirmed / contradicted / unknown, with the plain-words evidence that produced it) weighed against the request. A result may also carry place_note ("Near Williamsburg", "In Hoboken"): where that room is, present ONLY when the room is outside the place asked for — a room in the place the diner asked for carries no place_note, so never read a missing place_note as "somewhere else", and when one IS present say so rather than implying the room is in the place asked for. When result_groups is present, results is already partitioned into ORDERED, LABELED groups (exact / exact_within_wider_radius / likely_match_unconfirmed / nearby_any / relaxed_amenity) — present them to the user top to bottom, in that order, using each group's own message verbatim; never re-order the groups or the rows within a group (present rows in the order given), and never offer them as a choice. Never claim a nearby_any or relaxed_amenity row satisfies a constraint it does not carry in honored_constraints, and never call a likely_match_unconfirmed row (or any row carrying a requirement with status "unknown") a confirmed or exact match — say plainly that it likely matches but could not be confirmed from Seemor's data, and name what could not be confirmed (unknown_requirements lists it at the response level). repeat_ask: true means this is the same answer already given for the identical ask; say so rather than presenting it as new. Call it ONCE per diner ask, with the full ask (cuisine, occasion, vibe, constraints) in query and the place in location. Pass the diner's request in the diner's own words: do not add emphasis, priorities or qualities the diner did not state (for example "prioritize a quiet atmosphere" when the diner only said "quiet"), because added emphasis turns a light preference into a hard requirement and changes the answer. Results are ranked and final: do not re-call with reworded variations or call lookup_restaurant on each result to double-check it; that adds latency, not a better answer. For follow-ups about one place, use lookup_restaurant or ask_about_restaurant. If a result carries caveats, coverage_level 'basic', or a message noting a thin pool, relay it to the user instead of searching again. Requires a location: include one in your query (e.g. "in Soho"), or provide location, or latitude/longitude, or the tool refuses with location_required instead of guessing a city. Rows carry google_maps_url and website when Seemor holds them (null otherwise), plus seemor_url. Seemor holds no booking, availability or opening-hours data: use a booking app for those and join on restaurant name and address.
search_restaurants
ChatGPTSearch for restaurants near a place. Name the place with location (for example "Shibuya, Tokyo"), or give its latitude and longitude; one of the two is required, and location wins when both are sent. A name several places share returns status geo_conflict with candidates to choose from, and a name Seemor does not know returns status not_found. For picking where to eat or giving recommendations, use recommend; this tool is for browsing a list by distance or filters. Returns graded, ranked results with cuisine, price level, and Seemor analysis summaries. Fully analyzed restaurants (coverage_level 'full', letter grade) come first; when fewer than limit are available, quick-read restaurants are appended after them, marked coverage_level 'basic' with grade null and a preliminary_band (e.g. 'A-range') graded from review analysis, not star ratings; the band is the floor of the range, so 'B-range' means at least B minus. total_in_area = analyzed_in_area + preliminary_in_area; a message explains when quick-read rows are included. When location names a place Seemor holds, rooms inside that place come first and any room outside it carries a place_note (present it as outside the place asked for); analyzed_in_place counts the analyzed rooms inside and returned_outside_place counts the returned rooms outside.
ask_about_restaurant
ClaudeAsk a specific question about a restaurant based on analysis of real reviews and menu data. Common questions: what to order, group suitability, dietary options, vibe/atmosphere, value assessment. Requires a restaurant_id from find_restaurant or search_restaurants. Ask one question per call. Accounts connected through an assistant have a small free number of answers per day; if the result carries limit_reached, relay its message to the user as written and do not retry.
explore_area
ClaudeGet aggregate dining statistics for a neighborhood, city, or region: cuisine breakdown, grade distribution, price range, top neighborhoods, and highlighted restaurants. Use this for area-level context ("what is the dining scene like in Shoreditch?"), NOT for finding a specific restaurant or getting a personal recommendation. For "find me a quiet Italian near Shoreditch", use the recommend tool instead. Coverage is reported honestly: coverage_level 'full' (fully analyzed rows present), 'basic' (only quick-read rows: review-analysis bands, no letter grades yet), or 'none'; analyzed_count and preliminary_count split total_restaurants, and highlights marked coverage_level 'basic' carry a preliminary_band instead of a grade. not_found means neither analyzed nor preliminary rows exist.
find_restaurant
ClaudeFind restaurants by name. Returns matching restaurants with their Seemor IDs, which can be passed to lookup_restaurant for full details. Use the city parameter to disambiguate common names. Each result reports coverage_level: 'full' (letter grade), 'basic' (preliminary read: grade null, preliminary_band, tldr, caveat) or 'none' (no analysis yet).
lookup_restaurant
ClaudeLook up a specific restaurant by its Seemor ID. Returns grade, summary, cuisine, neighborhood, and other details. Use the fields parameter to request richer data (standard or premium; fully analyzed restaurants only). coverage_level 'full' rows carry a letter grade; 'basic' rows are Seemor quick reads: review-analysis bands (grade null, preliminary_band such as 'B-range') with a one-line tldr, graded from review analysis rather than star ratings; 'none' rows have no analysis yet. Use search_restaurants or find_restaurant first to get restaurant IDs. Use this for a single place the user asks about, not for every result of recommend.
recommend
ClaudeReturns a complete, ranked restaurant recommendation for a dining occasion: the final answer, not a search. Each result already carries a grade (or preliminary_band), match_reasons, caveats (a caveat beginning "Misses" names something the room does not deliver, never a good thing about it; a strong_match row never carries one), honored_constraints (plain-language labels for the stated filters this specific row confirms, e.g. "Indian", "Takeaway", "In Magnolia"), and requirements (per stated requirement, one of confirmed / contradicted / unknown, with the plain-words evidence that produced it) weighed against the request. A result may also carry place_note ("Near Williamsburg", "In Hoboken"): where that room is, present ONLY when the room is outside the place asked for — a room in the place the diner asked for carries no place_note, so never read a missing place_note as "somewhere else", and when one IS present say so rather than implying the room is in the place asked for. When result_groups is present, results is already partitioned into ORDERED, LABELED groups (exact / exact_within_wider_radius / likely_match_unconfirmed / nearby_any / relaxed_amenity) — present them to the user top to bottom, in that order, using each group's own message verbatim; never re-order the groups or the rows within a group (present rows in the order given), and never offer them as a choice. Never claim a nearby_any or relaxed_amenity row satisfies a constraint it does not carry in honored_constraints, and never call a likely_match_unconfirmed row (or any row carrying a requirement with status "unknown") a confirmed or exact match — say plainly that it likely matches but could not be confirmed from Seemor's data, and name what could not be confirmed (unknown_requirements lists it at the response level). repeat_ask: true means this is the same answer already given for the identical ask; say so rather than presenting it as new. Call it ONCE per diner ask, with the full ask (cuisine, occasion, vibe, constraints) in query and the place in location. Results are ranked and final: do not re-call with reworded variations or call lookup_restaurant on each result to double-check it; that adds latency, not a better answer. For follow-ups about one place, use lookup_restaurant or ask_about_restaurant. Accounts connected through an assistant have a small free number of recommend runs per day (only runs that return results count); if the result carries limit_reached, relay its message to the user as written and do not retry. If a result carries caveats, coverage_level 'basic', or a message noting a thin pool, relay it to the user instead of searching again. Requires a location: include one in your query (e.g. "in Soho"), or provide location, or latitude/longitude, or the tool refuses with location_required instead of guessing a city.
search_restaurants
ClaudeSearch for restaurants near a place. Name the place with location (for example "Shibuya, Tokyo"), or give its latitude and longitude; one of the two is required, and location wins when both are sent. A name several places share returns status geo_conflict with candidates to choose from, and a name Seemor does not know returns status not_found. For picking where to eat or giving recommendations, use recommend; this tool is for browsing a list by distance or filters. Returns graded, ranked results with cuisine, price level, and Seemor analysis summaries. Fully analyzed restaurants (coverage_level 'full', letter grade) come first; when fewer than limit are available, quick-read restaurants are appended after them, marked coverage_level 'basic' with grade null and a preliminary_band (e.g. 'A-range') graded from review analysis, not star ratings. total_in_area = analyzed_in_area + preliminary_in_area; a message explains when quick-read rows are included. When location names a place Seemor holds, rooms inside that place come first and any room outside it carries a place_note (present it as outside the place asked for); analyzed_in_place counts the analyzed rooms inside and returned_outside_place counts the returned rooms outside.