Fix: Normalze all database distances to score in [0, 1]

backend/open_webui/retrieval/vector/dbs/chroma.py

@@ -75,10 +75,16 @@ class ChromaClient:
                     n_results=limit,
                 )
 
+                # chromadb has cosine distance, 2 (worst) -> 0 (best). Re-odering to 0 -> 1
+                # https://docs.trychroma.com/docs/collections/configure cosine equation
+                distances: list = result["distances"][0]
+                distances = [2 - dist for dist in distances]
+                distances = [[dist/2 for dist in distances]]
+
                 return SearchResult(
                     **{
                         "ids": result["ids"],
-                        "distances": result["distances"],
+                        "distances": distances,
                         "documents": result["documents"],
                         "metadatas": result["metadatas"],
                     }

backend/open_webui/retrieval/vector/dbs/milvus.py

@@ -64,7 +64,10 @@ class MilvusClient:
 
             for item in match:
                 _ids.append(item.get("id"))
-                _distances.append(item.get("distance"))
+                # normalize milvus score from [-1, 1] to [0, 1] range
+                # https://milvus.io/docs/de/metric.md
+                _dist = (item.get("distance") + 1.0)/2.0
+                _distances.append(_dist)
                 _documents.append(item.get("entity", {}).get("data", {}).get("text"))
                 _metadatas.append(item.get("entity", {}).get("metadata"))
 

backend/open_webui/retrieval/vector/dbs/opensearch.py

@@ -120,7 +120,7 @@ class OpenSearchClient:
                     "script_score": {
                         "query": {"match_all": {}},
                         "script": {
-                            "source": "cosineSimilarity(params.query_value, doc[params.field]) + 1.0",
+                            "source": "(cosineSimilarity(params.query_value, doc[params.field]) + 1.0) / 2.0",
                             "params": {
                                 "field": "vector",
                                 "query_value": vectors[0],

backend/open_webui/retrieval/vector/dbs/pgvector.py

@@ -278,7 +278,9 @@ class PgvectorClient:
             for row in results:
                 qid = int(row.qid)
                 ids[qid].append(row.id)
-                distances[qid].append(row.distance)
+                # normalize and re-orders pgvec distance from [2, 0] to [0, 1] score range
+                # https://github.com/pgvector/pgvector?tab=readme-ov-file#querying
+                distances[qid].append((2.0 - row.distance)/2.0)
                 documents[qid].append(row.text)
                 metadatas[qid].append(row.vmetadata)
 

backend/open_webui/retrieval/vector/dbs/qdrant.py

@@ -99,7 +99,8 @@ class QdrantClient:
             ids=get_result.ids,
             documents=get_result.documents,
             metadatas=get_result.metadatas,
-            distances=[[point.score for point in query_response.points]],
+            # qdrant distance is [-1, 1], normalize to [0, 1]
+            distances=[[(point.score + 1.0)/2.0 for point in query_response.points]],
         )
 
     def query(self, collection_name: str, filter: dict, limit: Optional[int] = None):