feat: add local agent MCP tools and optional source-grounded AI answers

crates/searchi-ai/README.md

@@ -0,0 +1,44 @@
+# Optional source-grounded AI
+
+This GUI-free crate sends a bounded explicit request to a user-configured
+OpenAI-compatible Chat Completions endpoint. It is independent of search and
+contains no default provider, endpoint, model, implicit setup call or download.
+Validating configuration and constructing a client do not contact the endpoint.
+
+`AiConfig {base_url, model, api_key}` accepts an explicit HTTP(S) base such as
+`http://127.0.0.1:1234/v1`; `/chat/completions` is appended. URL credentials,
+queries and fragments are rejected. Credentials are optional, header-only,
+marked sensitive, and redacted by configuration Debug. Redirects and ambient
+proxies are disabled. Reqwest with rustls provides portable maintained HTTP/TLS.
+Errors contain static descriptions/status codes, never raw URLs, provider error
+bodies, credentials or selected text. The crate performs no logging.
+
+`AiClient::generate(&AiRequest)` is the sole network operation. Callers must
+require explicit enablement and a user action, validate that sources belong to
+the current index, and obtain consent before sending excerpts to the configured
+endpoint. Desktop code should use a blocking worker; background searches must
+never depend on this client. Session credential storage is a caller concern.
+
+Requests contain `AiMode::Summary` or `Question`, an optional question, and
+1–8 `AiSource` records with short unique IDs, title, text and optional location.
+Bounds: 16 KiB per source, 32 KiB total selected text, 2 KiB question, 96 KiB
+encoded HTTP payload, 64 KiB response and 16 KiB answer. Request timeout is
+45 seconds including response consumption, with a five-second connect limit.
+There is no automatic retry or silent truncation.
+
+Sources are explicitly untrusted data in the system instruction; source commands
+cannot authorize tools or application actions. The only Chat Completions fields
+sent are model and messages, for broad provider compatibility. The model must
+return a JSON answer/citations object inside `choices[0].message.content`.
+Returned IDs must exist, citations must appear inline as `[id]`, and the inline
+ID set must match the citation array. Unknown or inconsistent IDs are rejected.
+A no-citation answer remains an **uncited draft**; citation validation does not
+prove factual grounding or neutralize all model prompt injection. Callers must
+label output accordingly and retain the original source for verification.
+
+Local HTTP tests inspect actual endpoint path/messages/header-only credential,
+untrusted source boundaries, offline validation, pre-network input limits,
+malformed/unknown/mismatched citations, keyless local endpoints, actual timeout,
+response cap and error redaction. They use no real provider credentials, cloud
+endpoints, GPU inference or simulated production model output. Real-provider
+behavior and Linux/macOS execution remain untested release checks.