
The Bocha AI open platform
The awkward part of letting a question-answering app browse the web is messy results with no dates. The search call can request a summary, limit freshness, and set how many hits to return, and each hit can include both a short snippet and a longer summary for a prompt. If you already have candidate documents, the semantic reranker orders them by the question instead of keyword overlap alone. Tool-call results are shaped for a model to read, and output can be streamed or returned in one response. It fits developers adding outside knowledge to agents, chatbots, and retrieval apps. You need your own key; a sample command in the docs is not account credit, and prices live on a separate pricing page.
Bocha Wanxiang is the Bocha AI open platform. It offers a web search API and a semantic reranking API. Search accepts a natural-language question and returns titles, links, snippets, summaries, and publish times across news, images, video, encyclopedic pages, and academic sources. The reranker runs after keyword-plus-vector retrieval and orders results against the question. Developers call the official API, invoke it as a model tool, or install a plugin on an agent platform.
Difficulty: Intermediate
Natural-language web search
Search with a sentence and receive titles, links, site names, snippets, summaries, and publish times, with freshness and result count under your control.
Rerank after hybrid retrieval
After keyword-plus-vector retrieval, a semantic reranker orders the hits against the question so relevance is not keyword overlap alone.
Results shaped for models
When invoked as a tool, search results are formatted for a model to read and can be streamed or returned at once.
Plugins on agent platforms
Besides the official API, you can install the search plugin on an agent-building platform and attach outside knowledge to an existing agent.
Grounding a chatbot
After a user question, call search, place dated summaries into the prompt, then generate the answer.
Agents looking up sources
Install the plugin or use a tool call so an agent can look up news, encyclopedia pages, or academic sources for a question.
Reranking candidates you have
When the product already has a set of documents, send them to the reranker so order follows the question rather than keyword count.
The request can carry a natural-language question plus freshness, whether to include a summary, and how many hits. The response includes titles, links, site names, snippets, longer summaries, and publish times. The sample only shows the fields; a real call needs your own API key. It is for agents, chatbots, and retrieval apps, not a public search box without an account.
The search API fetches hits from content sources. If you already have candidate documents, or hybrid retrieval ordered them poorly, the semantic reranker can reorder them against the question. The page says that model runs after keyword-plus-vector retrieval. The two APIs can be used separately; a search does not have to be reranked every time.
The site lists three paths: call the API yourself, invoke it as a model tool, or install the plugin on an agent platform. Tool results are formatted for a model and can be streamed or returned at once. Every path uses a key on your account. Open the pricing page separately; the homepage sample does not include your quota.
Real reviews and feedback from users