github/langchain
1
0
Fork 0
mirror of https://github.com/hwchase17/langchain.git synced 2025-08-17 16:39:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(chroma): add missing Euclidean relevance-score function

Browse Source 
This patch restores relevance-scored retrieval for the [Chroma](cci:2://file:///d:/Github/langchain/libs/partners/chroma/langchain_chroma/vectorstores.py:143:0-1263:50) vector store.

Problem
`Chroma._select_relevance_score_fn()` defaults to the L2 (`"l2"`) metric, but the corresponding [_euclidean_relevance_score_fn()](cci:1://file:///d:/Github/langchain/libs/partners/chroma/langchain_chroma/vectorstores.py:755:4-763:44) was never implemented. Any call to [similarity_search_with_relevance_scores()](cci:1://file:///d:/Github/langchain/libs/core/langchain_core/vectorstores/base.py:533:4-580:36) therefore raised `AttributeError`, breaking default Chroma searches and the related test [test_chroma_with_relevance_score_custom_normalization_fn](cci:1://file:///d:/Github/langchain/libs/partners/chroma/tests/integration_tests/test_vectorstores.py:516:0-536:5).

Solution
Introduced `@staticmethod _euclidean_relevance_score_fn(distance: float) -> float` using the normalization `1 / (1 + distance)`, ensuring:
* distance = 0 → score = 1 (most relevant)
* distance → ∞ → score → 0 (least relevant)

Impact
• Re-enables relevance-score queries for Chroma with L2 distance.
• Unblocks dependent retrievers and integration tests.
• Keeps API behavior consistent with other vector stores (e.g., Qdrant).
This commit is contained in:
S. M. Mohiuddin Khan Shiam 2025-06-18 00:46:59 +06:00
parent bc1b5ffc91
commit a921c12494
1 changed files with 10 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
libs/partners/chroma/langchain_chroma/vectorstores.py
View File

@ -753,6 +753,16 @@ class Chroma(VectorStore):
return _results_to_docs_and_vectors(results)
@staticmethod
def _euclidean_relevance_score_fn(distance: float) -> float:
"""Normalize Euclidean (L2) distance to a [0, 1] relevance score.
Uses the transformation ``1 / (1 + distance)`` so that:
* distance == 0 -> score == 1 (most relevant)
* distance -> ∞ -> score -> 0 (least relevant)
"""
return 1.0 / (1.0 + float(distance))
def _select_relevance_score_fn(self) -> Callable[[float], float]:
"""Select the relevance score function based on collections distance metric.