feat(evidence): use server-side Qdrant BM25 retrieval

harness/tht/evidence/corpus/pipeline.py

@@ -123,7 +123,7 @@ class CorpusPipeline:
         self, *, store: CorpusStore, sources: list[EvidenceSource], embedder,
         vector_store: VectorStore, embedding_model: str, embedding_dimensions: int,
         chunk_policy: ChunkPolicy, pipeline_version: str, retain_published_generations: int = 3,
-        workspace_id: str | None = None,
+        workspace_id: str | None = None, sparse_language: str = "italian",
     ) -> None:
         self.store = store
         self.sources = sources
@@ -137,6 +137,9 @@ class CorpusPipeline:
             raise ValueError("retain_published_generations must be at least 1")
         self.retain_published_generations = retain_published_generations
         self.workspace_id = workspace_id
+        if sparse_language not in {"english", "italian"}:
+            raise ValueError("unsupported Qdrant BM25 language")
+        self.sparse_language = sparse_language
 
     def _assert_workspace_binding(self) -> None:
         manifest = self.store.active_manifest()
@@ -795,9 +798,8 @@ class CorpusPipeline:
             except Exception:
                 logger.debug("Failed to delete the unpublished vector generation", exc_info=True)
 
-    @staticmethod
     def _vector_record(
-        chunk: CanonicalChunk, embedding: list[float], generation: str, workspace_id: str,
+        self, chunk: CanonicalChunk, embedding: list[float], generation: str, workspace_id: str,
     ):
         record = VectorRecord(
             id=f"{workspace_id}:{generation}:{chunk.chunk_id}",
@@ -810,4 +812,10 @@ class CorpusPipeline:
                 "vector_generation": generation,
             },
         )
-        return VectorWriteRecord(record=record, embedding=embedding, content_hash=chunk.content_hash)
+        return VectorWriteRecord(
+            record=record,
+            embedding=embedding,
+            content_hash=chunk.content_hash,
+            sparse_text=chunk.content,
+            sparse_language=self.sparse_language,
+        )