Pesquisa e Web

Transcribe video/audio URL or local file to text + SRT using yt-dlp + Groq Whisper API (free tier). Preferred default for v2 transcription. Returns raw text and segments; summary is caller's responsibility.

Use for Chinese community forum discussions on Baidu Tieba when query involves Chinese tech topics, product feedback, or hobby communities not covered by xiaohongshu/zhihu.

Open-source deep research for AI agents: 40 channels, 10+ Chinese sources.

Explicit loop state for multi-round research — maintains gaps, visited URLs, bad URLs, failed-evaluator feedback, and decides whether to fire another round of queries or stop. Inspired by WebThinker's in-band search protocol, node-deepresearch's agent.ts state, and Scira's extreme-search.

Use for public Discourse-based community discussions, especially Linux DO, when queries need Chinese AI/dev forum posts and native operator troubleshooting.

Interactive browser automation (click, type, wait, screenshot, navigate) for dynamic pages that need user-like actions — pagination, login, form submission, infinite scroll. Uses Microsoft's official @playwright/mcp server; no autosearch-side Python. Use when fetch-jina and fetch-crawl4ai still cannot get the content.

Generate questions from 3-6 stakeholder / expert personas to expand coverage before searching. Complements decompose-task (which splits by task structure) by splitting by viewpoint. Borrowed from STORM's pre-writing research pattern — asking "what would an investor / a security researcher / an end-user / a maintainer" each want to know.

Use when compiling raw sources into the wiki, querying the knowledge base, running health checks or lint on the wiki, evolving or improving wiki coverage, or when user says 'compile', 'update the wiki', 'query', 'lint', 'health check', 'evolve', 'what's missing', or 'suggest improvements'. Also triggers on questions that should be answered from the knowledge base.

Represent a research plan as a directed graph — nodes are sub-questions, edges are "answer-depends-on" dependencies. Independent nodes can be executed in parallel; downstream nodes wait for their predecessors. Borrowed from MindSearch's WebSearchGraph pattern. Replaces list/tree decomposition for plans with non-linear dependencies.

Chinese-language podcasts searchable via the Apple iTunes store public API.

Append a single skill-execution event to the per-skill `experience/patterns.jsonl` file. Runs every time a leaf skill is used. Append-only — runtime AI never reads this file directly; only the compacted `experience.md` digest. Fast tier, no LLM required.

Transcribe video/audio URL or local file to text + SRT using yt-dlp + OpenAI Whisper API. Paid fallback for v2 transcription when Groq is rate-limited or unavailable. Returns raw text and segments; summary is caller's responsibility.