// The search: one collection, one query, cited answers. The embedding leaves the process (the
// procedure is an effect), the ranking is the vector store's, and every answer carries where it
// came from: the file, the chunk, and the snippet a person reads.

import { forbidden, internal } from "@conusai-kit/core"
import { MAX_SEARCH_RESULTS, type SearchResult } from "../contract/index.js"
import { requireCollection } from "./collections.js"
import { type KnowledgeDeps, type KnowledgeScope, knowledgeEnabled } from "./ports.js"

export async function searchKnowledge(
  deps: KnowledgeDeps,
  scope: KnowledgeScope,
  input: { collectionName: string; query: string; limit?: number | undefined },
): Promise<{ results: SearchResult[] }> {
  if (!(await knowledgeEnabled(deps, scope))) throw forbidden("knowledge is switched off for this project")
  await requireCollection(deps, scope, input.collectionName)
  const service = deps.search
  if (!service) throw internal("the knowledge search service is not bound in this process")
  const limit = Math.min(input.limit ?? 10, MAX_SEARCH_RESULTS)
  const hits = await service.search(scope, {
    collectionName: input.collectionName,
    query: input.query,
    limit,
    actorId: scope.actor.userId,
  })
  return {
    results: hits.map((hit): SearchResult => {
      return {
        fileId: hit.fileId,
        filename: hit.filename,
        collectionName: hit.collectionName,
        snippet: hit.snippet,
        score: hit.score,
        chunk: hit.chunk,
      }
    }),
  }
}