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

harness/tht/evidence/search.py

@@ -3,6 +3,7 @@
 import re
 
 from tht.evidence.corpus.store import CorpusStore
+from tht.ports.vector import VectorStoreError
 
 
 class CorpusWorkspaceMismatchError(RuntimeError):
@@ -12,12 +13,27 @@ class CorpusWorkspaceMismatchError(RuntimeError):
 class ActiveEvidenceSearcher:
     """Searcher facade that enforces ACTIVE generation predicates before LIMIT."""
 
-    def __init__(self, corpus: CorpusStore, delegate, expected_workspace_id: str | None = None):
+    def __init__(
+        self,
+        corpus: CorpusStore,
+        delegate,
+        expected_workspace_id: str | None = None,
+        evidence_language: str = "italian",
+    ):
         self.corpus = corpus
         self.delegate = delegate
         self.expected_workspace_id = expected_workspace_id
+        self.evidence_language = evidence_language
 
-    def search(self, embedding, top_n=10, kinds=None, metadata_filter=None):
+    def search(
+        self,
+        embedding,
+        top_n=10,
+        kinds=None,
+        metadata_filter=None,
+        query_text=None,
+        query_language=None,
+    ):
         requested = set(kinds) if kinds is not None else {
             "schema_table", "schema_column", "evidence", "memory", "solved_question",
         }
@@ -59,9 +75,13 @@ class ActiveEvidenceSearcher:
                         generation = mapping.get(document.document_id, manifest.vector_generation)
                         if generation:
                             by_generation.setdefault(generation, []).append(document.document_id)
+                    if by_generation and (not isinstance(query_text, str) or query_text.strip() == ""):
+                        raise VectorStoreError("Evidence hybrid query text is required")
                     for generation, document_ids in sorted(by_generation.items()):
                         hits.extend(self.delegate.search(
                             embedding, top_n=top_n, kinds=["evidence"],
+                            query_text=query_text,
+                            query_language=query_language or self.evidence_language,
                             metadata_filter={
                                 "vector_generation": generation,
                                 "document_ids": sorted(document_ids),
@@ -73,7 +93,11 @@ class ActiveEvidenceSearcher:
 
 def active_searcher(cfg, delegate, *, workspace_id: str | None = None):
     corpus_root = cfg.paths.artifacts.parent / "corpus"
-    return ActiveEvidenceSearcher(CorpusStore(corpus_root), delegate, workspace_id)
+    languages = {"en": "english", "it": "italian"}
+    language = languages.get(getattr(cfg, "language", "en"))
+    if language is None:
+        raise VectorStoreError("workspace language is unsupported for Qdrant BM25")
+    return ActiveEvidenceSearcher(CorpusStore(corpus_root), delegate, workspace_id, language)
 
 
 def validate_corpus_workspace(cfg, workspace_id: str) -> None: