Search
Principles
- Read local material first, then decide whether external search is needed.
- Priority: official docs > official repo > original paper > high-quality implementation > community experience.
- Tool split:
websearch/Exa for discovery, webfetch for page bodies, curl for raw text/headers/APIs.
- For GitHub repo implementation, issues, PRs, commits, releases: use
github-search explicitly. Generic web search is not enough.
- For models/datasets/weights: check GitHub + Hugging Face + ModelScope in parallel.
Multi-Axis Search Matrix
Open at least these dimensions in parallel:
- Direct problem search — the task as stated.
- Lower-level/decomposed search — split into concepts, subproblems, signals, targets, eval settings.
- Related search — equivalent wording, nearby tasks, alternate descriptions.
- Heuristic-rewrite search — synonyms, reverse questions, restatements, failure modes.
- Trend-platform search — scan
huggingface.co/papers/, alphaxiv.org, paperdigest.org/arxiv/ for recent work.
Do NOT search the user's original sentence once and stop.
Recovery Chain
On search failure (429, captcha, timeout, anti-bot):
- Switch to another engine or platform.
- Look up public retry interval /
Retry-After for the failed entry.
- Use
minimum interval + random jitter sleep, then decide whether to go back.
- Keep collecting from other sources in parallel.
Identity/Tag Graph Expansion
When search hits a person, repo, model, org, or tag: keep tracing newer work, latest release/commit/paper, same-tag pages. Do not stop on the first hit. If a newer work in the same network covers the task better, promote it.
Platform Rules
If the task involves a platform/contest/benchmark host, also search: submit quota, daily limits, cooldown, eval delay, public/private leaderboard, code/resource limits, submit format. Do not enter platform submission before rules are clear.
Output
- Search matrix with axes covered
- Rebuilt search question set
- Evidence with source, time, version
- Separate facts from inference
Have we collected enough? — explicit yes/no with reasons- If not enough:
still need more collection + next-round actions
1---2name: search3description: External evidence collection — web search, GitHub/HF/ModelScope. Multi-axis search with recovery chain.4---567# Search89## Principles1011- Read local material first, then decide whether external search is needed.12- Priority: official docs > official repo > original paper > high-quality implementation > community experience.13- Tool split: `websearch`/Exa for discovery, `webfetch` for page bodies, `curl` for raw text/headers/APIs.14- For GitHub repo implementation, issues, PRs, commits, releases: use `github-search` explicitly. Generic web search is not enough.15- For models/datasets/weights: check GitHub + Hugging Face + ModelScope in parallel.1617## Multi-Axis Search Matrix1819Open at least these dimensions in parallel:20211. **Direct problem search** — the task as stated.222. **Lower-level/decomposed search** — split into concepts, subproblems, signals, targets, eval settings.233. **Related search** — equivalent wording, nearby tasks, alternate descriptions.244. **Heuristic-rewrite search** — synonyms, reverse questions, restatements, failure modes.255. **Trend-platform search** — scan `huggingface.co/papers/`, `alphaxiv.org`, `paperdigest.org/arxiv/` for recent work.2627Do NOT search the user's original sentence once and stop.2829## Recovery Chain3031On search failure (429, captcha, timeout, anti-bot):321. Switch to another engine or platform.332. Look up public retry interval / `Retry-After` for the failed entry.343. Use `minimum interval + random jitter` sleep, then decide whether to go back.354. Keep collecting from other sources in parallel.3637## Identity/Tag Graph Expansion3839When search hits a person, repo, model, org, or tag: keep tracing newer work, latest release/commit/paper, same-tag pages. Do not stop on the first hit. If a newer work in the same network covers the task better, promote it.4041## Platform Rules4243If the task involves a platform/contest/benchmark host, also search: submit quota, daily limits, cooldown, eval delay, public/private leaderboard, code/resource limits, submit format. Do not enter platform submission before rules are clear.4445## Output4647- Search matrix with axes covered48- Rebuilt search question set49- Evidence with source, time, version50- Separate facts from inference51- `Have we collected enough?` — explicit yes/no with reasons52- If not enough: `still need more collection` + next-round actions
ztxtech