Web Search

Web Search gives agents and applications a project-scoped search capability over pluggable engines. Operators create instances under Data → Web Search — an instance is a named search engine (Bing, Brave, Serper, Tavily, self-hosted SearxNG, or keyless DuckDuckGo) with its own credentials, defaults, playground, usage stats and per-instance search logs. A project can hold several instances, e.g. a keyless DuckDuckGo for development next to a Brave instance for production traffic.

The landing page tracks instance counts, active status and how many instances have AI answers enabled. Rows link to the instance detail page.

Engines

Driver	Type	Credentials	Notes
bing	Commercial	API key	Bing Web Search API v7; Azure regional endpoint override, market (mkt) setting
brave-search	Commercial	API key	Brave's independent index; country + language settings
serper	Commercial	API key	Google SERP via serper.dev; organic results + answer box
tavily	Commercial	API key	LLM-optimized search; searchDepth, provider-native answers
searxng	Open source, self-hosted	optional HTTP basic auth	Point at your instance's base URL (JSON format must be enabled); engine list configurable. Outbound requests go through the SSRF guard
duckduckgo	Keyless	none	Best-effort parse of the DuckDuckGo HTML endpoint; suitable for light usage

All engines share common settings (language, country/market, safe search) and normalize results to one shape: title / url / snippet / position plus publishedAt, source and score where available.

Creating an instance

Create instance opens the service picker filtered to web search engines; the driver's form schema (API key, base URL, safe search, …) renders dynamically. Credentials are encrypted at rest and never returned to the UI.

Instances are strictly project-scoped: they are created in the active project and are not visible from other projects.

Instance detail

The detail page follows the standard tabbed layout:

Playground

Run live queries through the instance; every run — from the API or this page — is recorded in Logs. The AI answer switch requests an AI interpretation of the results (see below).

Usage

Aggregates recent runs into search counts, success rate, error count, and average / p95 latency, plus a breakdown by caller source (api vs dashboard) and AI-answered searches.

Logs

Every search with its query, result count, latency, source and status. Filter by free text (queries, results and answers are searched), date range, and status. Clicking a row opens the full detail — the returned results and the synthesized answer are stored with each log entry.

Configuration

Shows the engine settings summary (edit re-opens the driver form) and hosts the AI Answer card and the danger zone.

AI answers

Like Tavily's synthesized answers, any instance can interpret its search results with a model. Enable it per instance under Configuration → AI Answer: toggle the switch, pick an LLM from the project's Model Hub, and optionally add extra instructions (tone, output format).

Requests then opt in per call with include_answer: true; the top results are handed to the model, which answers the query citing result numbers ([1], [2]). The response carries answer and answer_model.

If a request asks for an answer while the instance has AI answers disabled (or no model selected), the request fails with a clear error before the search provider is called.

API

Token-authenticated endpoints live under /api/client/v1/websearch/*:

# Search on a named instance
curl -X POST https://console.example.com/api/client/v1/websearch/brave-main/search \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "query": "fastify v5 changes", "count": 5, "include_answer": true }'

When the project has exactly one active instance the generic endpoint may omit the instance key; with multiple instances the request must name one (via the provider field or the /:key/search path).

See the Web Search API reference and the Console SDK (client.webSearch.search(...), client.webSearch.searchWith(key, ...)).

Access control

Web Search is a first-class RBAC service (websearch): grant read for listing/logs and write for creating instances and running searches. Instance credentials are stored on the shared provider record infrastructure (AES-256-GCM, tenant-scoped).