Answers

Answers a question from one or more collections, with citations.#

POST/lmkit/v1/search/answer

The grounded-answer endpoint: retrieves the best-matching content from the selected collections (every retrieval control of the search endpoint applies: search_type, filters, rerank, query_mode, context expansion), generates an answer on a local chat model bound to those sources, and returns the answer with citations. Citations cover EVERY retrieved source: 'cited' marks the ones the answer references inline as [n], and semantic matches carry the exact character span of the supporting chunk within the document's stored page Markdown. 'grounding' controls strictness: 'Strict' (default) answers from the sources only and refuses a question that retrieves nothing without calling the model; 'Relaxed' lets clearly-flagged general knowledge fill gaps. Set 'stream' for Server-Sent Events (delta frames, a terminal frame with the full result, then 'data: [DONE]'). Search type degrades automatically: 'Hybrid' falls back to whichever mode the tenant has enabled.

Request body

application/json ·

Property	Type	Description
`cluster_id`	string	The search cluster that holds the tenant and collections. If not provided, the default cluster is used.
`tenant_id`required	string (uuid)	The tenant that owns the collections.
`collection_id`	string (uuid)	A single collection to answer from. Ignored when 'collection_ids' is provided.
`collection_ids`	string (uuid)[]	The set of collections to answer from in one call. Takes precedence over 'collection_id'.
`question`required	string	The question to answer.
`search_type`		How supporting sources are retrieved: 'Hybrid' (default) combines word and meaning matching and degrades automatically to whichever mode the tenant has enabled; 'Semantic' and 'FullText' force one mode.
`fusion`		Hybrid retrieval only: how the full-text and semantic arms are combined. 'Convex' (default) blends normalized scores, preserving each arm's confidence; 'Rrf' fuses by rank.
`groundedness`		Verifies the generated answer against the sources it was given, which catches the case a score floor cannot: retrieval returned plausible same-topic passages that do not answer the question. 'Off' (default) skips the check; 'Report' returns the answer unchanged with its supported-claim fraction and the unsupported claims; 'Refuse' replaces an answer below 'groundedness_threshold' with the grounded refusal. The check costs additional model passes.
`groundedness_threshold`	object (double)	Supported-claim fraction at or above which an answer is kept under 'Refuse'. Defaults to 1.
`top_k`	object (int32)	How many sources to retrieve and ground the answer on (1..20). Defaults to 5.
`filters`	object	Optional metadata filter (equality map), as on the search endpoint. Mutually exclusive with 'filter'.
`filter`	one of object ·
`min_score`	object (double)	Optional relevance floor for retrieved sources, as on the search endpoint. Defaults to 0.
`rerank`	boolean	When true, a cross-encoder reranking pass re-scores the retrieved sources before grounding, as on the search endpoint.
`rerank_top_n`	object (int32)	The reranking over-fetch depth, as on the search endpoint. Defaults to 50.
`rerank_alpha`	object (double)	The rerank blend, as on the search endpoint. Defaults to 1.
`query_mode`		Query understanding, as on the search endpoint: 'Contextual' (with 'chat_history'), 'MultiQuery', or 'Hyde' (semantic/hybrid retrieval only). Defaults to 'Original'.
`chat_history`	[]	Prior conversation turns for contextual query rewriting, as on the search endpoint.
`query_variants`	object (int32)	MultiQuery variant count (1..5), as on the search endpoint. Defaults to 3.
`context_expansion`		How much surrounding content each source contributes: 'Chunk' (default: the matched chunk with its neighbors; full pages for full-text matches) or 'Page' (the whole matched page).
`max_chunks_per_document`	object (int32)	Maximum sources drawn from one document (1..10), so one strong document does not crowd out the rest. Defaults to 3.
`model`	string	The chat model that generates the answer. Empty uses the server's default chat model.
`grounding`		How strictly the answer is bound to the sources. 'Strict' (default): answer from the sources only; a question the sources cannot answer is declined, and one that retrieves nothing is refused without an inference pass ('answer' is empty and 'context_found' false). 'Relaxed': general knowledge may fill gaps, flagged as such.
`prompt_template`	string	Optional custom prompt template. Must contain '@context' (replaced by the numbered source blocks) and '@question'. Empty uses the built-in grounded template.
`max_completion_tokens`	object (int32)	Maximum tokens the answer may generate. Defaults to 1024.
`temperature`	object (double)	Sampling temperature. 0 (default) decodes greedily, which suits grounded answering; above 0 samples.
`max_context_characters`	object (int32)	Character budget for the source blocks in the prompt (sources past the budget are returned as retrieved-but-unused). Defaults to 24000.
`stream`	boolean	When true, the answer streams as Server-Sent Events: data frames carrying 'delta' (and 'thinking' for reasoning models), a terminal frame carrying the full answer with 'citations' and 'done': true, then the literal 'data: [DONE]'.

Responses

Status	Type	Description
200		OK
400		Bad Request
404		Not Found
503	application/json	Service Unavailable

curl -X POST "$LMKIT_ONE_URL/lmkit/v1/search/answer" \
  -H "Authorization: Bearer $LMKIT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "cluster_id": "string",
  "tenant_id": "string",
  "collection_id": "string",
  "collection_ids": [
    "string"
  ],
  "question": "",
  "search_type": "Hybrid"
}'