Port v5-revert params from Lotus (best score 0.5517 vs our 0.5094)

Score formula: recall×0.8 + ndcg×0.2 → recall 4x more important

index: CHUNK_SIZE 256→384 (sweet spot, not too small, not too large)
search: DENSE 80→50, SPARSE 200→150, RETRIEVE 150→100, RERANK 10→15
  Fewer candidates = less noise = better recall

Lotus experiments confirmed: 80/200/150 limits HURT vs 50/150/100.

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

index/main.py

@@ -74,7 +74,7 @@ class SparseVector(BaseModel):
     values: list[float]
 
 
-CHUNK_SIZE = 256
+CHUNK_SIZE = 384
 OVERLAP_SIZE = 128
 SPARSE_MODEL_NAME = "Qdrant/bm25"
 FASTEMBED_CACHE_PATH = "/models/fastembed"

search/main.py

@@ -161,10 +161,10 @@ async def lifespan(app: FastAPI):
 
 app = FastAPI(title="Search Service", version="0.1.0", lifespan=lifespan)
 
-DENSE_PREFETCH_K = 80
-SPARSE_PREFETCH_K = 200
-RETRIEVE_K = 150
-RERANK_LIMIT = 10
+DENSE_PREFETCH_K = 50
+SPARSE_PREFETCH_K = 150
+RETRIEVE_K = 100
+RERANK_LIMIT = 15
 
 
 async def embed_dense(client: httpx.AsyncClient, text: str) -> list[float]: