Merge pull request #896 from khoj-ai/features/add-support-for-custom-…

…confidence

Add support for custom search model-specific thresholds

src/khoj/database/migrations/0059_searchmodelconfig_bi_encoder_confidence_threshold.py

@@ -0,0 +1,17 @@
+# Generated by Django 5.0.7 on 2024-08-24 18:19
+
+from django.db import migrations, models
+
+
+class Migration(migrations.Migration):
+    dependencies = [
+        ("database", "0058_alter_chatmodeloptions_chat_model"),
+    ]
+
+    operations = [
+        migrations.AddField(
+            model_name="searchmodelconfig",
+            name="bi_encoder_confidence_threshold",
+            field=models.FloatField(default=0.18),
+        ),
+    ]

src/khoj/database/models/__init__.py

@@ -270,6 +270,8 @@ class ModelType(models.TextChoices):
     cross_encoder_inference_endpoint = models.CharField(max_length=200, default=None, null=True, blank=True)
     # Inference server API Key to use for embeddings inference. Cross-encoder model should be hosted on this server
     cross_encoder_inference_endpoint_api_key = models.CharField(max_length=200, default=None, null=True, blank=True)
+    # The confidence threshold of the bi_encoder model to consider the embeddings as relevant
+    bi_encoder_confidence_threshold = models.FloatField(default=0.18)
 
 
 class TextToImageModelConfig(BaseModel):

src/khoj/routers/api.py

@@ -83,7 +83,7 @@ async def search(
         n=n,
         t=t,
         r=r,
-        max_distance=max_distance,
+        max_distance=max_distance or math.inf,
         dedupe=dedupe,
     )
 
@@ -117,7 +117,6 @@ async def execute_search(
     # initialize variables
     user_query = q.strip()
     results_count = n or 5
-    max_distance = max_distance or math.inf
     search_futures: List[concurrent.futures.Future] = []
 
     # return cached results, if available

src/khoj/routers/api_chat.py

@@ -524,7 +524,7 @@ async def chat(
     common: CommonQueryParams,
     q: str,
     n: int = 7,
-    d: float = 0.18,
+    d: float = None,
     stream: Optional[bool] = False,
     title: Optional[str] = None,
     conversation_id: Optional[int] = None,
@@ -764,7 +764,7 @@ def collect_telemetry():
             meta_log,
             q,
             (n or 7),
-            (d or 0.18),
+            d,
             conversation_id,
             conversation_commands,
             location,

src/khoj/search_type/text_search.py

@@ -100,18 +100,23 @@ async def query(
     raw_query: str,
     type: SearchType = SearchType.All,
     question_embedding: Union[torch.Tensor, None] = None,
-    max_distance: float = math.inf,
+    max_distance: float = None,
 ) -> Tuple[List[dict], List[Entry]]:
     "Search for entries that answer the query"
 
     file_type = search_type_to_embeddings_type[type.value]
 
     query = raw_query
+    search_model = await sync_to_async(get_user_search_model_or_default)(user)
+    if not max_distance:
+        if search_model.bi_encoder_confidence_threshold:
+            max_distance = search_model.bi_encoder_confidence_threshold
+        else:
+            max_distance = math.inf
 
     # Encode the query using the bi-encoder
     if question_embedding is None:
         with timer("Query Encode Time", logger, state.device):
-            search_model = await sync_to_async(get_user_search_model_or_default)(user)
             question_embedding = state.embeddings_model[search_model.name].embed_query(query)
 
     # Find relevant entries for the query