Skip to content

armgit / LEANN

Public Imported from GitHub
0

[MLsys2026 Best Paper]: https://arxiv.org/abs/2506.08276. RAG on Everything with LEANN. Enjoy 97% storage savings while running a fast, accurate, and 100% private RAG application on your personal device.

fix(core): drop the build template from query embeddings on the fallback path (#442)

Serhii Zghamadb2609817e7b
diff --git a/packages/leann-core/src/leann/searcher_base.py b/packages/leann-core/src/leann/searcher_base.py
index 0d22d40..57c059f 100644
--- a/packages/leann-core/src/leann/searcher_base.py
+++ b/packages/leann-core/src/leann/searcher_base.py
@@ -213,6 +213,15 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
self.zmq_connection = ReusableZMQConnection()
self._zmq_port: Optional[int] = None
+ def _search_provider_options(self) -> dict[str, Any]:
+ # compute_query_embedding applies the query template itself; a build template
+ # left in here would be prepended to the query by the provider as well.
+ return {
+ k: v
+ for k, v in self.embedding_options.items()
+ if k not in ("build_prompt_template", "query_prompt_template", "prompt_template")
+ }
+
def _load_meta(self) -> dict[str, Any]:
"""Loads the metadata file associated with the index."""
# This is the corrected logic for finding the meta file.
@@ -239,14 +248,7 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
or "mips"
)
- # Filter out ALL prompt templates from provider_options during search
- # Templates are applied in compute_query_embedding (line 109-110) BEFORE server call
- # The server should never apply templates during search to avoid double-templating
- search_provider_options = {
- k: v
- for k, v in self.embedding_options.items()
- if k not in ("build_prompt_template", "query_prompt_template", "prompt_template")
- }
+ search_provider_options = self._search_provider_options()
server_started, actual_port = self.embedding_server_manager.start_server(
port=port if port is not None else 5557,
@@ -335,7 +337,7 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
[query],
self.embedding_model,
embedding_mode,
- provider_options=self.embedding_options,
+ provider_options=self._search_provider_options(),
)
# Cache the result (use original query before template)
diff --git a/tests/test_prompt_template_persistence.py b/tests/test_prompt_template_persistence.py
index 6bf3761..b285f4b 100644
--- a/tests/test_prompt_template_persistence.py
+++ b/tests/test_prompt_template_persistence.py
@@ -896,3 +896,70 @@ class TestQueryTemplateApplicationInComputeEmbedding:
# Empty string is falsy, so no template should be applied
assert captured_queries[0] == "vector database"
+
+ @pytest.mark.parametrize(
+ "embedding_options, query_template",
+ [
+ (
+ {
+ "build_prompt_template": "title: none | text: ",
+ "query_prompt_template": "task: search result | query: ",
+ },
+ "task: search result | query: ",
+ ),
+ ({"prompt_template": "search_query: "}, "search_query: "),
+ ],
+ )
+ def test_fallback_path_does_not_add_build_template(
+ self, temp_index_with_template, embedding_options, query_template
+ ):
+ """The provider must see only the query template, not the build template on top."""
+ from leann.searcher_base import BaseSearcher
+
+ class TestSearcher(BaseSearcher):
+ def search(
+ self,
+ query,
+ top_k,
+ complexity=64,
+ beam_width=1,
+ prune_ratio=0.0,
+ recompute_embeddings=False,
+ pruning_strategy="global",
+ zmq_port=None,
+ **kwargs,
+ ):
+ return {"labels": [], "distances": []}
+
+ searcher = object.__new__(TestSearcher)
+ searcher.index_path = Path(temp_index_with_template)
+ searcher.index_dir = searcher.index_path.parent
+ with open(searcher.index_dir / f"{searcher.index_path.name}.meta.json") as f:
+ searcher.meta = json.load(f)
+ searcher.embedding_model = searcher.meta["embedding_model"]
+ searcher.embedding_mode = searcher.meta["embedding_mode"]
+ searcher.embedding_options = {"api_key": "test-key", **embedding_options}
+ searcher.meta["embedding_options"] = searcher.embedding_options
+ searcher.query_cache = Mock()
+ searcher.query_cache.get.return_value = None
+
+ sent_inputs = []
+
+ def create(model, input, **kwargs):
+ sent_inputs.extend(input)
+ return Mock(data=[Mock(embedding=[0.1] * 768) for _ in input])
+
+ client = Mock()
+ client.embeddings.create.side_effect = create
+
+ with (
+ patch("openai.OpenAI", return_value=client),
+ patch("leann.embedding_compute.get_model_token_limit", return_value=8192),
+ ):
+ searcher.compute_query_embedding(
+ query="vector database",
+ use_server_if_available=False,
+ query_template=query_template,
+ )
+
+ assert sent_inputs == [f"{query_template}vector database"]