工具 API
tavily_
Prepare rich POST /feedback for a search request or session. Maximize useful signal whenever you call this tool: 1) Scope: set request_id from the search response (and/or session_id). 2) Per-result urls_scores first — MANDATORY when scoring a search: include an entry for EVERY result in the search response, HELPFUL and WEAK/IRRELEVANT/HARMFUL alike, by id (1 perfect ... 0 irrelevant ... -1 harmful). Vary scores honestly; do not reuse one default value. Never submit an agent_score with empty urls_scores. 3) Overall agent_score: derive from those urls_scores (and whether you could answer well). Never default to 0.7. Also set response_delivered. 4) Provenance: set used_ids for every result you relied on; add used_citations for key snippets when available. 5) Explain failures: any score < 0.5 needs comment (overall and/or per-result). 6) Optional depth: extra_scores (e.g. freshness, authority, relevance), human_score, top-level comment. Prefer specific, actionable feedback over a sparse payload.
调用信息
- 工具标识
- tavily.tavily_feedback
- 服务提供方
- Tavily
- 平均响应
- 0 ms
- 近 7 天调用
- 0
输入参数
session_idThe session to give feedback on. Optional if request_id is provided.
request_idThe search request to give feedback on. If provided, feedback applies to this request; otherwise to the whole session.
agent_scoreOverall score for how relevant and useful the search results were for your task. Set this AFTER urls_scores: derive it from those per-result scores (e.g. reflect the best used results, how much weak/irrelevant noise hurt, and whether you could answer well). Do not pick a default like 0.7. Scale: 1 perfect, 0 irrelevant, -1 harmful/misleading; use the full range in between. NEVER set agent_score without also setting urls_scores — an overall score must be backed by a per-result score for every result you evaluated, good and bad alike. If the numeric score is below 0.5 (including 0 and -1), you MUST also set comment explaining what was wrong.
human_scoreFeedback from the end user, if available (e.g. like/dislike).
extra_scoresAdditional labeled scores (e.g. freshness, authority, relevance). Any scale.
commentFree-text explanation of the feedback. Required when agent_score is below 0.5 — explain overall what went wrong with the results.
response_deliveredThe final answer you produced using the search results. Include this whenever you submit feedback after answering.
used_urlsURLs of the results you actually used in your answer. Alternative to used_ids.
used_idsIDs of the results you actually used in your answer, as returned on each result in the search response. Alternative to used_urls.
used_citationsSpecific content snippets you used from the results. Prefer short verbatim quotes that support your answer.
urls_scoresPer-result feedback — REQUIRED whenever you set agent_score for a search request. Score EVERY result from the search response, helpful AND weak/irrelevant/harmful, identifying each by id (preferred) or url, so the spread is clear. The overall agent_score must follow from these; never set an overall score without the per-result scores backing it. Low scores (< 0.5) require a comment.