Ask — AskEngine

Ask the belief layer a whole question and get one synthesized, cited answer. Synchronous and slower than the read endpoints — one LLM completion per call, so expect seconds.

Every citation is verified against the signals and evidence the model actually saw — a ref or evidence id it invented never reaches the wire. Check a cited belief with get signal and fetch the cited quotes with get evidence via evidenceIds. When the material is thin, the answer says the belief layer does not cover the question instead of guessing.

POST/v1/ask

Example

curl -X POST "https://silent-viper-157.convex.site/v1/ask" \  -H "Authorization: Bearer $ASKENGINE_API_KEY" \  -H "Content-Type: application/json" \  -d '{    "question": "What do users actually think about onboarding since the redesign?"  }'

Request Body

questionstring

One focused question in plain language. Doubles as the semantic-search input, so it shares search's cap.

Response

{  "data": {    "answer": "The evidence points to onboarding confusion concentrating on API-key setup (SIG-4): \"I gave up twice before I figured out where to add my API key.\" — Ada L., study \"Onboarding interviews Q4\". Counter-evidence is thin — only one participant called the flow easy — so the belief stands but is not unanimous.",    "citations": [      {        "ref": "SIG-4",        "signalId": "sig_123",        "evidenceIds": [          "ev_001",          "ev_007"        ]      }    ],    "searchedAt": 1764453600000  }}