Fix search reliability: batch dense embedding, graceful extra-query fallback, rerank 429 retry

- embed_dense_multi now sends one batch request (N texts → 1 API call) instead of N parallel
  requests, avoiding rate-limit errors when question has variants/hyde
- Extra dense embeddings (variants/hyde) wrapped in try/except so primary query always succeeds
- Reranker now retries up to 5 times with exponential backoff on 429, matching Lotus reference

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

search/main.py

@@ -80,16 +80,18 @@ async def search(payload: SearchAPIRequest) -> SearchAPIResponse:
     extra_texts = build_extra_dense_queries(question)
     sparse_text = build_sparse_query(question)
 
-    async def _no_extra() -> list:
-        return []
-
-    extra_task = embed_dense_multi(client, extra_texts) if extra_texts else _no_extra()
-    primary_dense, extra_dense_vecs, sparse_vec = await asyncio.gather(
+    primary_dense, sparse_vec = await asyncio.gather(
         _embed_dense_with_retry(client, primary_query),
-        extra_task,
         asyncio.to_thread(embed_sparse, sparse_text),
     )
 
+    extra_dense_vecs: list[list[float]] = []
+    if extra_texts:
+        try:
+            extra_dense_vecs = await embed_dense_multi(client, extra_texts[:3])
+        except Exception as exc:
+            logger.warning("Extra dense embedding failed, continuing without it: %s", exc)
+
     points = await qdrant_search(
         qdrant,
         primary_dense,

search/query_builder.py

@@ -38,9 +38,24 @@ async def embed_dense(client: httpx.AsyncClient, text: str) -> list[float]:
     return payload.data[0].embedding
 
 
+async def embed_dense_batch(client: httpx.AsyncClient, texts: list[str]) -> list[list[float]]:
+    """Single batch request for multiple texts (avoids N parallel calls)."""
+    response = await client.post(
+        str(EMBEDDINGS_DENSE_URL),
+        **get_upstream_kwargs(),
+        json={
+            "model": os.getenv("EMBEDDINGS_DENSE_MODEL", EMBEDDINGS_DENSE_MODEL),
+            "input": texts,
+        },
+    )
+    response.raise_for_status()
+    payload = DenseEmbeddingResponse.model_validate(response.json())
+    payload.data.sort(key=lambda x: x.index)
+    return [item.embedding for item in payload.data]
+
+
 async def embed_dense_multi(client: httpx.AsyncClient, texts: list[str]) -> list[list[float]]:
-    tasks = [embed_dense(client, t) for t in texts]
-    return list(await asyncio.gather(*tasks))
+    return await embed_dense_batch(client, texts)
 
 
 def embed_sparse(text: str) -> SparseVector:

search/rerank.py

@@ -14,6 +14,10 @@ async def get_rerank_scores(
     if not targets:
         return []
 
+    import asyncio as _asyncio
+
+    for attempt in range(5):
+        try:
             response = await client.post(
                 str(RERANKER_URL),
                 **get_upstream_kwargs(),
@@ -24,10 +28,21 @@ async def get_rerank_scores(
                     "text_2": targets,
                 },
             )
+            if response.status_code == 429:
+                wait = 2 ** attempt
+                logger.warning("Rerank 429, retry %d/5 in %ds", attempt + 1, wait)
+                await _asyncio.sleep(wait)
+                continue
             response.raise_for_status()
-
             data = response.json().get("data") or []
             return [float(sample["score"]) for sample in data]
+        except Exception as exc:
+            if attempt < 4:
+                await _asyncio.sleep(2 ** attempt)
+                continue
+            raise exc
+
+    return []
 
 
 async def rerank_points(