Inside Yuanbao Search: how AI search for agents hits sub-300ms with GistToken caching
赛博禅心 · wechat · 2026-10-21
The author, an early builder of ChatGPT's built-in search plugin, breaks down Tencent Yuanbao's new AI search service and how search is being re-engineered for agents:
- Core loop: the model only decides what to search and how to use results; the search side handles quality and ranking
- Agent-friendly signals: search and agent exchange authority scores, citations and answer ratings, cutting searches per task from 8 to 5 in joint evals
- Quality control: three-level site/account/page scoring, dedicated models against AIGC spam and GEO content
- Freshness: 157 top data sources across 13 verticals via plugin cards (stocks, flights, etc.)
- Latency: hybrid inverted+vector index over hundreds of billions of pages (94% recall, 8ms); pages pre-compressed into 100 GistTokens cached in KVCache, saving 30% inference cost; sub-300ms per query
- Integration: Tencent Cloud API (¥20/1k calls, 2,000 free/month), plus MCP and Skill options; a Codex demo shows three-hop Q&A, with search as atomic capability and hop planning left to the agent
More from coding & agent
- Index Network open-sources Social Discovery Protocol: agents negotiate fit on your behalf — carsonfarmer · 2026-10-23
- Installed a formal-verification skill: it silently installed Java, MacBook runs 10 degrees hotter — zealcaiden · 2026-10-23
- OpenAI's Decisions API hits GA: near real-time routing up to 10x faster than GPT-6 Luna — OpenAIDevs · 2026-10-23
- Ruff author warns: upgrade Node/Bun now — --use-system-ca still accepts the eMudhra cert — charliermarsh · 2026-10-23
- Proof of Agent: paste one prompt and your AI mines, grades, and self-governs — adamamcbride · 2026-10-23
- DeepLearning.AI weekly: GPT-6.1 hits 97.9% byte parity on COBOL migration — DeepLearningAI · 2026-10-23