Implement Phase 12 feedback, Phase 13 Semantic Chunker, Phase 13-B Reranker, Bug 5 thinking fix

- Phase 12: FeedbackRepository + td_feedback 테이블, Gradio 👍/👎 이벤트, run_id 추적, LangSmith create_feedback() 연동
- Phase 13: 커스텀 _SemanticSplitter 제거 → langchain_experimental.SemanticChunker 교체, buffer_size/threshold_type 환경변수 적용
- Phase 13-B: RerankService (Cross-Encoder), RetrieverService.search()에 reranker 통합, tools.py as_retriever() → search() 전환
- Bug 5: mlx_chat_model enable_thinking 런타임 오버라이드, agent_service stream_mode=["messages","custom"] 이중 스트림, thinking 토큰 custom 이벤트로 emit
- ROADMAP: LLM 모델명 8B 반영, RAG에 Reranker 추가, 추천 진행 순서 갱신

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

services/rag/ingestion_service.py

@@ -1,59 +1,10 @@
-import re
-
-import numpy as np
 from langchain_community.document_loaders import PDFPlumberLoader, TextLoader
-from langchain_core.documents import Document
+from langchain_experimental.text_splitter import SemanticChunker
 from langchain_qdrant import QdrantVectorStore
 from qdrant_client import QdrantClient
 from qdrant_client.models import Filter, FieldCondition, MatchValue, FilterSelector
 
 
-def _cosine_similarity(a: np.ndarray, b: np.ndarray) -> float:
-    return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b) + 1e-10))
-
-
-class _SemanticSplitter:
-    """문장 임베딩 유사도 기반 청커.
-
-    인접 문장 간 코사인 유사도를 계산하고, 유사도가 낮은(= 의미 전환) 지점에서 청크를 분리한다.
-    breakpoint_percentile=95이면 유사도 하위 5% 지점이 분리 경계가 된다.
-    """
-
-    _SENTENCE_RE = re.compile(r"(?<=[.!?。！？])\s+")
-
-    def __init__(self, embeddings, breakpoint_percentile: int = 95):
-        self._embeddings = embeddings
-        self._percentile = breakpoint_percentile
-
-    def split_documents(self, docs: list[Document]) -> list[Document]:
-        result = []
-        for doc in docs:
-            for chunk_text in self._split_text(doc.page_content):
-                result.append(Document(page_content=chunk_text, metadata=doc.metadata))
-        return result
-
-    def _split_text(self, text: str) -> list[str]:
-        sentences = [s for s in self._SENTENCE_RE.split(text.strip()) if s.strip()]
-        if len(sentences) <= 1:
-            return [text.strip()] if text.strip() else []
-
-        vecs = np.array(self._embeddings.embed_documents(sentences))
-        similarities = [_cosine_similarity(vecs[i], vecs[i + 1]) for i in range(len(vecs) - 1)]
-        threshold = float(np.percentile(similarities, 100 - self._percentile))
-        breakpoints = [i + 1 for i, s in enumerate(similarities) if s < threshold]
-
-        chunks, start = [], 0
-        for bp in breakpoints:
-            chunk = " ".join(sentences[start:bp]).strip()
-            if chunk:
-                chunks.append(chunk)
-            start = bp
-        tail = " ".join(sentences[start:]).strip()
-        if tail:
-            chunks.append(tail)
-        return chunks
-
-
 class IngestionService:
     """문서를 의미 단위 청크로 분할해 Qdrant에 저장하는 수집 파이프라인."""
 
@@ -63,12 +14,16 @@ class IngestionService:
         qdrant_url: str,
         collection_name: str,
         breakpoint_threshold_type: str = "percentile",
+        buffer_size: int = 1,
     ):
         self._embeddings = embeddings
         self._qdrant_url = qdrant_url
         self._collection_name = collection_name
-        # breakpoint_threshold_type은 향후 확장용으로 수용 (현재는 percentile 방식 고정)
-        self._splitter = _SemanticSplitter(embeddings, breakpoint_percentile=95)
+        self._splitter = SemanticChunker(
+            embeddings=embeddings,
+            breakpoint_threshold_type=breakpoint_threshold_type,
+            buffer_size=buffer_size,
+        )
         self._client = QdrantClient(url=qdrant_url)
 
     def _delete_by_source(self, source_path: str) -> None: