feat: Add rerank score filtering with configurable threshold

- Add DEFAULT_MIN_RERANK_SCORE constant (default: 0.0)
- Add MIN_RERANK_SCORE environment variable support
- Filter chunks with rerank scores below threshold in process_chunks_unified
- Add info-level logging for filtering operations
- Handle empty results gracefully after filtering
- Maintain backward compatibility with non-reranked chunks

env.example

@@ -67,6 +67,9 @@ ENABLE_LLM_CACHE=true
 
 ### Reranker configuration (Set ENABLE_RERANK to true in reranking model is configed)
 # ENABLE_RERANK=True
+### Minimum rerank score for document chunk exclusion (set to 0.0 to keep all chunks, 0.6 or above if LLM is not strong enought)
+# MIN_RERANK_SCORE=0.0
+### Rerank model configuration (required when ENABLE_RERANK=True)
 # RERANK_MODEL=jina-reranker-v2-base-multilingual
 # RERANK_BINDING_HOST=https://api.jina.ai/v1/rerank
 # RERANK_BINDING_API_KEY=your_rerank_api_key_here

lightrag/api/config.py

@@ -19,6 +19,7 @@ from lightrag.constants import (
     DEFAULT_MAX_TOTAL_TOKENS,
     DEFAULT_COSINE_THRESHOLD,
     DEFAULT_RELATED_CHUNK_NUMBER,
+    DEFAULT_MIN_RERANK_SCORE,
 )
 
 # use the .env that is inside the current folder
@@ -290,6 +291,11 @@ def parse_args() -> argparse.Namespace:
     args.rerank_binding_host = get_env_value("RERANK_BINDING_HOST", None)
     args.rerank_binding_api_key = get_env_value("RERANK_BINDING_API_KEY", None)
 
+    # Min rerank score configuration
+    args.min_rerank_score = get_env_value(
+        "MIN_RERANK_SCORE", DEFAULT_MIN_RERANK_SCORE, float
+    )
+
     # Query configuration
     args.history_turns = get_env_value("HISTORY_TURNS", DEFAULT_HISTORY_TURNS, int)
     args.top_k = get_env_value("TOP_K", DEFAULT_TOP_K, int)

lightrag/constants.py

@@ -23,6 +23,9 @@ DEFAULT_ENABLE_RERANK = True
 DEFAULT_COSINE_THRESHOLD = 0.2
 DEFAULT_RELATED_CHUNK_NUMBER = 5
 
+# Rerank configuration defaults
+DEFAULT_MIN_RERANK_SCORE = 0.0
+
 # Separator for graph fields
 GRAPH_FIELD_SEP = "<SEP>"
 

lightrag/lightrag.py

@@ -31,6 +31,7 @@ from lightrag.constants import (
     DEFAULT_MAX_TOTAL_TOKENS,
     DEFAULT_COSINE_THRESHOLD,
     DEFAULT_RELATED_CHUNK_NUMBER,
+    DEFAULT_MIN_RERANK_SCORE,
 )
 from lightrag.utils import get_env_value
 
@@ -284,6 +285,11 @@ class LightRAG:
     rerank_model_func: Callable[..., object] | None = field(default=None)
     """Function for reranking retrieved documents. All rerank configurations (model name, API keys, top_k, etc.) should be included in this function. Optional."""
 
+    min_rerank_score: float = field(
+        default=get_env_value("MIN_RERANK_SCORE", DEFAULT_MIN_RERANK_SCORE, float)
+    )
+    """Minimum rerank score threshold for filtering chunks after reranking."""
+
     # Storage
     # ---
 

lightrag/utils.py

@@ -1871,7 +1871,29 @@ async def process_chunks_unified(
             top_n=rerank_top_k,
         )
 
-    # 2. Apply chunk_top_k limiting if specified
+    # 2. Filter by minimum rerank score if reranking is enabled
+    if query_param.enable_rerank and unique_chunks:
+        min_rerank_score = global_config.get("min_rerank_score", 0.5)
+        original_count = len(unique_chunks)
+
+        # Filter chunks with score below threshold
+        filtered_chunks = []
+        for chunk in unique_chunks:
+            rerank_score = chunk.get("rerank_score", 1.0)  # Default to 1.0 if no score
+            if rerank_score >= min_rerank_score:
+                filtered_chunks.append(chunk)
+
+        unique_chunks = filtered_chunks
+        filtered_count = original_count - len(unique_chunks)
+
+        if filtered_count > 0:
+            logger.info(
+                f"Rerank filtering remained: {len(unique_chunks)} chunks (min rerank score: {min_rerank_score})"
+            )
+        if not unique_chunks:
+            return []
+
+    # 3. Apply chunk_top_k limiting if specified
     if query_param.chunk_top_k is not None and query_param.chunk_top_k > 0:
         if len(unique_chunks) > query_param.chunk_top_k:
             unique_chunks = unique_chunks[: query_param.chunk_top_k]
@@ -1879,7 +1901,7 @@ async def process_chunks_unified(
             f"Kept chunk_top-k: {len(unique_chunks)} chunks (deduplicated original: {origin_count})"
         )
 
-    # 3. Token-based final truncation
+    # 4. Token-based final truncation
     tokenizer = global_config.get("tokenizer")
     if tokenizer and unique_chunks:
         # Set default chunk_token_limit if not provided