What it does
Hybrid retrieval + reranking + LLM synthesis. Returns a complete answer with inline [N] citations and the source chunks they map to.
Required scope
answers:read
Input
| Field | Type | Required | Default | Description |
|---|---|---|---|---|
query |
string | yes | — | The question, in natural language |
scope |
enum | no | "all" |
"all", "files", "meetings", "calls" |
timeRange |
enum | no | "any" |
"any", "today", "week", "month" |
Output
{
"answer": "SHINE AI maintains SOC 2 Type II attestation covering Security, Availability, and Confidentiality [1]. The auditor is Prescient Assurance LLP [1].",
"citations": [
{
"fileId": "de2e3b54-388b-47d1-a27a-445af0a7df02",
"fileName": "soc2-controls.md",
"chunkIndex": 0,
"snippet": "SHINE AI maintains a SOC 2 Type II attestation..."
}
],
"usage": {
"tokens_in": 1842,
"tokens_out": 196
}
}Example invocation
A user asks:
What did Brian quote Trimark?
The AI app calls shine_ask({ query: "What did Brian quote Trimark?" }) and gets back a grounded answer with citations to the pricing-conversations file.
Latency
Typical: 2-6 seconds end-to-end. LLM synthesis dominates (~2-5s). Retrieval itself is ~500ms.
Notes
- The answer is retrieval-augmented. If the corpus doesn’t cover the question, the answer will honestly say so rather than hallucinate.
- Citations are stable — the AI app can present them as source links back to the SHINE web UI.