Autosearch

Chinese tech business news, startup funding, and industry analysis from 36kr.

Transcribe video/audio URL to text + word-level timestamps using Bilibili Bcut ASR API (free, no API key). Preferred for Chinese content — Bcut gives character-level timestamps vs Whisper word-level. Returns text + segments [{start, end, text}]. Requires yt-dlp + ffmpeg.

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.

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.

Transcribe video/audio URL or local file to text + SRT using yt-dlp + local mlx-whisper (Apple Silicon). Free, offline, fastest on M-series Macs. Opt-in advanced path for users with Apple Silicon + mlx-whisper installed. Returns raw text and segments; summary is caller's responsibility.

Deep URL fetch using crawl4ai (Playwright-powered) for JS-rendered pages, anti-bot sites, and dynamic content. Slower than fetch-jina but handles sites that block simple fetchers. Requires user-installed crawl4ai package.

Use to scrape a URL into clean Markdown when the page is JS-rendered, behind anti-bot, or a PDF — and FIRECRAWL_API_KEY is available. Degrades gracefully to warn when key is missing.

Fetch any public URL as clean Markdown via Jina Reader (r.jina.ai). Best for articles, docs, blog posts; weak on heavy anti-bot sites (Zhihu, Xiaohongshu).

Promote recurring patterns from `experience/patterns.jsonl` into the compact `experience.md` digest (≤120 lines, read by runtime AI before calling the skill). Triggers on N-events / file-size / user-feedback / session-end. Guards against single-success noise and pollution via promotion thresholds.

Advisory skill — three-tier (Fast / Standard / Best) model routing catalog for autosearch skills. Tells the runtime AI which tier each leaf skill needs, and how to escalate or de-escalate. Autosearch does not switch models itself; the runtime AI is the decision-maker.

Route an autosearch task to the smallest relevant group of skills, then load only the matching leaf skills. First line of defense against loading all 80+ skill SKILL.md files at session start.

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.

Route the runtime AI to free third-party MCP servers via the upstream `mcporter` router — semantic web search (Exa), Chinese UGC (Weibo / Douyin / Xiaohongshu), and professional (LinkedIn). Zero API keys; free fallback path before reaching for paid TikHub on hard Chinese anti-bot platforms.

Group-first channel selection algorithm for v2 tool-supplier architecture. Given a research query + clarify rubrics + channel_priority hints, picks 1-3 relevant groups from the router index, then 3-8 leaf channels from within those groups. Replaces flat-rank selection across 41 channels with a two-stage pick so runtime AI never has to read all 41 SKILL.md bodies.

Current news headlines aggregated across publishers via Google News RSS (English US feed).

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.

Instagram public posts and reels search via TikHub. Best for product demos, tutorials, and visual content. English-first but works with any language keyword.

WeChat Channels (视频号) short video search via TikHub. Returns videos matching a keyword — creator name, title, duration, CDN video URL. Best for Chinese-language video content on WeChat ecosystem. Requires TIKHUB_API_KEY.

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

Chinese stock market platform — stock search, trending posts, hot stock rankings. Use for financial/investment queries in Chinese.

Chinese developer community discussions on programming, tech, and career — useful for native-zh developer sentiment and niche technical troubleshooting.

LinkedIn public pages — company profiles, job posts, professional articles via Jina Reader (no auth needed for public content).

DuckDuckGo Search — free general web search with no auth, use as broad default for any English or mixed query.

Use for Docker image and container registry searches when query involves Docker images, containers, or deployment artifacts.

Discover open machine learning models on Hugging Face Hub via the public model search API.

Chinese engineering articles covering architecture, AI, and enterprise tech from InfoQ 中文, via public RSS feed.

Search scholarly works through OpenAlex's public works search API with open-access URL fallback.

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

Use for biomedical and life-science literature searches when query expects PubMed articles, clinical trials, or medical research papers.

Reddit community discussions, user experience reports, and topic debates via the public search.json endpoint.

Use for general web meta-search via a local SearXNG instance when broad multi-engine web coverage is needed and SEARXNG_URL is configured.

US public company filings (10-K, 10-Q, 8-K) via SEC EDGAR full-text search — financial and regulatory disclosures for research.

Chinese WeChat Official Account articles via the public Sogou WeChat search SERP.

Programming Q&A with community-voted answers across 200+ technical tags via api.stackexchange.com.

Structured entity data (people, places, concepts) from Wikidata knowledge graph.

Authoritative encyclopedia articles via the Wikipedia Action API (English edition).

Deduplicate URLs across all sources, assign stable citation numbers, and merge citations from multiple subagents / sections into one consistent reference list. Prevents "same URL cited as [3] in one paragraph and [17] in another" and "different URLs merged under [5]" bugs that come from per-section synthesis.

Session-level policy for keeping the runtime AI's context window healthy across long research — keep-last-k tool results, offload older evidence to disk, trigger compaction at thresholds. Borrows MiroThinker's keep_tool_result, deepagents' summarization middleware, and deer-flow's SummarizationEvent pattern. Orthogonal to assemble-context (which is per-synthesis-step); this is per-session.

Define the execution contract for isolating a research sub-task — input schema, budget, return summary, evidence list, failure status. Complements decompose-task (which only splits the problem) by giving each split a bounded, auditable execution unit the runtime AI can farm out to a sub-agent or parallel session.

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.