zhizhi-search

--- name: zhizhi-search description: Use when the user needs a specific answer from the web and that answer can be judged right or wrong — fact-checking, source hunting, and looking up details all count. Also use after one or more rounds of searching when results start drifting away from what the user actually asked. Do not use for casual chat, pure reasoning, writing, or when existing material already answers. license: MIT version: 1.3.0 --- # Zhizhi · Drift Checkpoint for Search **Core principle:** This is not a search method. It is a judgment you must pass after every round of searching — is what you are holding still the thing the user asked for? ## When to Use - The user needs an answer that can be judged right or wrong, and you are about to search the web. - You have searched one or more rounds and the results no longer match the user's original wording. - You are about to rephrase the user's question and search again. **Not for:** casual chat, pure reasoning, writing, or when existing material already answers. ## Core Pattern **Before** (default behavior): > The user asks where A comes from → you cannot find it → you swap A for a near-synonym B → you find plenty about B → you deliver an explanation of B. The user's subject was replaced, you answered a different question, and you never said so. **After**: > Round 1 searches A and returns only B → judge: drifted → stop and report "found B, not A itself" → end with one question asking whether to change direction. ## Workflow Four steps per question, in order. Each step names its input and its output; do not start the next step until the current one has produced its output. 1. **Capture the subject.** Input: the user's exact wording. Output: the main subject, copied verbatim — the thing the question is *about*, not the thing you expect to find. 2. **Search the subject as written.** Input: the subject from step 1. Output: this round's results. 3. 🔴 **CHECKPOINT — judge.** Input: the subject from step 1 plus this round's results. Output: exactly one verdict, `Matched`, `Drifted`, or `Not found`, plus one line saying why. - `Matched` → go to step 4. - `Drifted` or `Not found` → STOP searching, go to step 4. Do not return to step 2 under a new subject. 4. **Deliver.** Input: the verdict from step 3. Output: the delivery format that matches that verdict. Reporting "I could not find it" is a complete result, not a failure of the workflow. Reaching step 4 from a `Drifted` or `Not found` verdict is success. ## When a Round Fails | Trigger | First fix | If it still fails | |---|---|---| | Zero results for the subject | Search it again as a quoted phrase, then in the user's own language | Deliver `Not found`; name what you searched and where you stopped | | Every hit is about a near-synonym, not the subject | Judge `Drifted` before reading further | Report what you found, say why it is not the subject, ask one question | | The source your conclusion depends on will not open | Try an archived copy of the page | Mark the claim second-hand or unopened; do not promote it to a conclusion | | Two sources disagree on the fact | Open both and compare their dates | Report both values with dates; never pick one silently | | The user replies "no, I meant X" after a drift report | That is permission to re-target: X is now the subject | Search X once; do not resume the abandoned subject | ## Implementation ### Hard constraints (violating any one voids the search) 1. **Do not change the main subject.** The main subject in the user's own wording must not be replaced. You may suggest alternatives, but a suggestion must be explicitly labeled as a suggestion and must not replace the user's direction during the search. Check: was the main subject swapped for another word or another object? 2. **Judge every round.** After each round of searching there must be one judgment — no skipping, no merging. Check: did each round leave a judgment? 3. **Do not re-search a judged path.** A path already judged "drifted" must not be walked again in an equivalent form. Check: was the same rejected path launched a second time? 4. **Say "not found" when it is not found.** Never turn "I cannot find it" into "they probably meant another word". Check: did you report not-found as found, or substitute something else? ### Soft constraints (optional, but state in the delivery why you skipped them) - Label second-hand sources as second-hand; open the evidence your conclusion depends on, and if you cannot open it, say so. - Write `unknown` when you cannot judge; never "probably". - Stop after three equivalent rounds with no new evidence. ### Delivery - **Matched:** one conclusion + one quotable, independently reproducible piece of evidence + source. - **Drifted:** state what you actually found and where it differs from the request; end with one question (change direction / narrow it down) — one question, no follow-up, no answering for the user. - **Not found:** state what you searched and where you stopped; end with the same one question. ## Do Not Do This | Mistake | Fix | |---|---| | Cannot find it, so swap the subject for a synonym and search again | Judge "drifted", stop and report; if a swap is warranted, ask first | | Treat "suggest changing direction" as the new direction and just search it | Label it as a suggestion and wait for the user | | Judge "drifted", then reword and walk the same path | Do not launch a rejected path again | | Turn not-found into "they probably meant X" and answer X | Say not found; stay not found | | Read a whole page about the near-synonym "just to check" | Judge first; reading the wrong page is how the subject slips |

存档时间线

版本索引永久保留；内容副本只保留最新 1 份，更早版本仅留索引与哈希（存档时间线的证据链不会因此断裂）。

下载存档内容副本