feat(rag): optimize RAG pipeline — JSON-Mode, CoT, Hybrid Search, Re-Ranking, Cross-Reg Dedup, chunk 1024

Phase 1 (LLM Quality):
- Add format=json to all Ollama payloads (obligation_extractor, control_generator, citation_backfill)
- Add Chain-of-Thought analysis steps to Pass 0a/0b system prompts

Phase 2 (Retrieval Quality):
- Hybrid search via Qdrant Query API with RRF fusion + automatic text index (legal_rag.go)
- Fallback to dense-only search if Query API unavailable
- Cross-encoder re-ranking with BGE Reranker v2 (RERANK_ENABLED=false by default)
- CPU-only PyTorch dependency to keep Docker image small

Phase 3 (Data Layer):
- Cross-regulation dedup pass (threshold 0.95) links controls across regulations
- DedupResult.link_type field distinguishes dedup_merge vs cross_regulation
- Chunk size defaults updated 512/50 → 1024/128 for new ingestions only
- Existing collections and controls are NOT affected

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend-compliance/compliance/services/reranker.py

@@ -0,0 +1,85 @@
+"""
+Cross-Encoder Re-Ranking for RAG Search Results.
+
+Uses BGE Reranker v2 (BAAI/bge-reranker-v2-m3, MIT license) to re-rank
+search results from Qdrant for improved retrieval quality.
+
+Lazy-loads the model on first use. Disabled by default (RERANK_ENABLED=false).
+"""
+
+import logging
+import os
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+RERANK_ENABLED = os.getenv("RERANK_ENABLED", "false").lower() == "true"
+RERANK_MODEL = os.getenv("RERANK_MODEL", "BAAI/bge-reranker-v2-m3")
+
+
+class Reranker:
+    """Cross-encoder reranker using sentence-transformers."""
+
+    def __init__(self, model_name: str = RERANK_MODEL):
+        self._model = None  # Lazy init
+        self._model_name = model_name
+
+    def _ensure_model(self) -> None:
+        """Load model on first use."""
+        if self._model is not None:
+            return
+        try:
+            from sentence_transformers import CrossEncoder
+
+            logger.info("Loading reranker model: %s", self._model_name)
+            self._model = CrossEncoder(self._model_name)
+            logger.info("Reranker model loaded successfully")
+        except ImportError:
+            logger.error(
+                "sentence-transformers not installed. "
+                "Install with: pip install sentence-transformers"
+            )
+            raise
+        except Exception as e:
+            logger.error("Failed to load reranker model: %s", e)
+            raise
+
+    def rerank(
+        self, query: str, texts: list[str], top_k: int = 5
+    ) -> list[int]:
+        """
+        Return indices of top_k texts sorted by relevance (highest first).
+
+        Args:
+            query: The search query.
+            texts: List of candidate texts to re-rank.
+            top_k: Number of top results to return.
+
+        Returns:
+            List of indices into the original texts list, sorted by relevance.
+        """
+        if not texts:
+            return []
+
+        self._ensure_model()
+
+        pairs = [[query, text] for text in texts]
+        scores = self._model.predict(pairs)
+
+        # Sort by score descending, return indices
+        ranked = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)
+        return ranked[:top_k]
+
+
+# Module-level singleton
+_reranker: Optional[Reranker] = None
+
+
+def get_reranker() -> Optional[Reranker]:
+    """Get the shared reranker instance. Returns None if disabled."""
+    global _reranker
+    if not RERANK_ENABLED:
+        return None
+    if _reranker is None:
+        _reranker = Reranker()
+    return _reranker