Agent Skills › tavily-ai/skills › tavily-dynamic-search

tavily-dynamic-search

GitHub

针对需过滤、去重或多步筛选的大规模网页搜索结果,通过本地CLI和脚本将原始内容落盘并提取关键证据,避免大模型上下文溢出。

skills/tavily-dynamic-search/SKILL.md tavily-ai/skills

触发场景

需要处理大量网页搜索结果的过滤与去重 网页内容过大需先保存再提取关键信息以避免上下文溢出

安装

npx skills add tavily-ai/skills --skill tavily-dynamic-search -g -y
更多选项

不安装直接使用

npx skills use tavily-ai/skills@tavily-dynamic-search

指定 Agent (Claude Code)

npx skills add tavily-ai/skills --skill tavily-dynamic-search -a claude-code -g -y

安装 repo 全部 skill

npx skills add tavily-ai/skills --all -g -y

预览 repo 内 skill

npx skills add tavily-ai/skills --list

SKILL.md

Frontmatter
{
    "name": "tavily-dynamic-search",
    "description": "Programmatic Tavily search with context isolation. Use when web search or extraction will return large results that need filtering, deduplication, or multi-step triage before they enter the model context. Use tavily-search for ordinary lookups and tavily-research for end-to-end cited synthesis.\n",
    "allowed-tools": "Bash(tvly *), Bash(python3 *), Bash(uv run *), Bash(jq *)"
}

Tavily Dynamic Search

Keep large raw web payloads on disk and return only the evidence needed for the task. This is useful when using --include-raw-content, combining several queries, or extracting multiple long pages. Do not use this workflow for a simple lookup that a normal tvly search --json can answer directly.

Before running

Search and extract support capped keyless access. Run them directly when tvly is available. If tvly is missing, follow the tavily-cli setup. Do not look for an API key or authenticate before the first request. If the keyless cap is reached in an interactive session, run tvly login to open browser OAuth, then retry the blocked request once. In an unattended environment, report the cap and authentication options instead of starting an interactive flow.

Workflow

  1. Search broadly without raw content and inspect titles, URLs, scores, and snippets.
  2. Fetch full content only for the best sources.
  3. When raw output could be large, save it with -o and filter the file before printing anything to the model context.
  4. Preserve source URLs beside every extracted fact.

When the user restricts evidence to official or named domains, validate the hostname of every selected URL during local filtering. --include-domains narrows the search but is not proof that every returned result belongs to an allowed host. If full-page extraction is unavailable, label conclusions as search-snippet evidence instead of implying that the page body was verified.

Keep the process in one turn when the relevant sources and filters are already known. Use another turn only when the first search changes what should be extracted.

Create a unique temporary task directory before saving evidence so concurrent agents do not overwrite one another. Python's tempfile.mkdtemp() is available when mktemp is not permitted. Reuse that directory for all raw and filtered artifacts from the task.

Small result: filter a direct JSON response

For a small search response, a pipe is enough:

tvly search "query" --max-results 5 --json | python3 -c '
import json, sys
data = json.load(sys.stdin)
for result in data.get("results", []):
    score = result.get("score") or 0
    title = result.get("title") or ""
    print(f"[{score:.2f}] {title}")
    print(result.get("url", ""))
    print(result.get("content", "")[:300])
'

Do not discard stderr. Authentication failures, keyless-cap messages, and API errors are actionable and must remain visible.

Large result: save first, then filter

Use the CLI's file output so raw page content does not pass through the tool response:

tvly search "query" \
  --include-raw-content markdown \
  --max-results 8 \
  --json \
  -o /tmp/tavily-search-results.json

Then print only bounded evidence:

python3 -c '
import json
from pathlib import Path

data = json.loads(Path("/tmp/tavily-search-results.json").read_text())
for result in data.get("results", []):
    title = result.get("title") or ""
    url = result.get("url") or ""
    print(f"## {title}")
    print(f"URL: {url}")
    print((result.get("raw_content") or result.get("content") or "")[:1200])
    print()
'

Adjust the filtering logic to the question. Prefer relevant paragraphs or fields over fixed character slices when the target information is known. Aim for roughly 150-600 tokens per source unless a table or code block genuinely requires more.

Targeted extraction

When search identifies the right URLs, extract only those pages:

tvly extract "https://example.com/article" \
  --json \
  -o /tmp/tavily-extract-results.json

For topic-focused pages, let Tavily reduce the response before local filtering:

tvly extract "https://example.com/docs" \
  --query "authentication API" \
  --chunks-per-source 3 \
  --json \
  -o /tmp/tavily-extract-results.json

Multiple queries

For multi-angle research, run a small set of focused searches, deduplicate by URL, and rank before extracting. Use subprocess.run(..., capture_output=True, text=True) when orchestrating commands in Python. Check returncode; if a command fails, surface its stderr and stop or retry deliberately. Never use a blanket except Exception: continue that hides missing evidence.

Response shapes

tvly search --json returns query, optional answer, results, and response_time. Each result commonly contains url, title, content, score, and optional raw_content.

tvly extract --json returns results, failed_results, and response_time. Each successful result commonly contains url, raw_content, and optional images.

Treat fields as optional and use .get() while filtering. Inspect failed_results instead of assuming every requested URL succeeded.

Useful options

Option Purpose
--max-results Bound the search result count; default 5, maximum 20
--depth Choose ultra-fast, fast, basic, or advanced
--time-range Restrict results to day, week, month, or year
--include-domains Restrict results to a comma-separated list of trusted domains
--exclude-domains Exclude a comma-separated list of domains
--include-raw-content Include full content as markdown or text
-o, --output Save the complete response to a file

Use jq only for short filters when Python is unavailable:

tvly search "query" --json | jq '[.results[] | {title, url, score, content}]'

版本历史

  • 778122e 当前 2026-09-09 08:45

    文档更新:优先配置免密钥访问;强化技能工作流规范;对齐引导式CLI设置说明。

  • ea5e820 2026-07-25 07:40

同 Skill 集合

skills/tasks/tavily-entity/SKILL.md
skills/tavily-best-practices/SKILL.md
skills/tavily-cli/SKILL.md
skills/tavily-extract/SKILL.md
skills/tavily-map/SKILL.md
skills/tavily-research/SKILL.md
skills/core/tavily-search/SKILL.md
skills/tasks/tavily-list/SKILL.md
skills/tasks/tavily-monitor/SKILL.md
skills/tasks/tavily-research/SKILL.md
skills/tavily-crawl/SKILL.md
skills/tavily-search/SKILL.md

元信息

文件数
0
版本
778122e
Hash
3c443efa
收录时间
2026-07-25 07:40

首页 - Wiki
Copyright © 2011-2026 iteam. Current version is 2.156.0. UTC+08:00, 2026-11-18 15:27
浙ICP备14020137号-1