改进rerank效果

backend/.env.xinference

@@ -3,6 +3,16 @@
 SQL_DATABASE_URL=mysql+pymysql://zjinfo1:Dy2Bcr53Hm5xRkba@110.42.234.166:3306/zjinfo1
 #SQL_DATABASE_URL=mysql+pymysql://zjinfo2:GSKcziSdBixDXwcd@110.42.234.166:3306/zjinfo2
 
+#--------------------------
+# 是否启用检索重排功能
+ENABLE_RERANK=true
+# The number of similar embeddings to return when retrieving documents.
+TOP_K=5
+# Rerank model
+RERANK_MODEL=bge-reranker-v2-m3
+RERANK_BASE_URL=http://10.1.16.39:9995
+RERANK_TOP_N=5
+RERANK_THRESHOLD=0.3
 #----------   Xinference    ----------------
 # The provider for the AI models to use.
 MODEL_PROVIDER=xinference
@@ -19,9 +29,7 @@ EMBEDDING_MODEL=bge-m3
 EMBEDDING_BASE_URL=http://10.1.16.39:9995
 # Dimension of the embedding model to use.
 EMBEDDING_DIM=1024
-# Rerank model
+
-RERANK_MODEL=bge-reranker-v2-m3
-RERANK_BASE_URL=http://10.1.16.39:9995
 ##----------   OpenAI    ----------------
 ## The provider for the AI models to use.
 #MODEL_PROVIDER=openai
@@ -46,17 +54,10 @@ RERANK_BASE_URL=http://10.1.16.39:9995
 ## Name of the embedding model to use.
 #EMBEDDING_MODEL=text-embedding-v2
 
-#--------------------------
-# 是否启用检索重排功能
-ENABLE_RERANK=true
-
 
 # The questions to help users get started (multi-line).
 CONVERSATION_STARTERS=本工程指什么？\n总算表有哪些费用？\n项目划分哪些内容构成？\n其他费用表有哪些内容？
 
-# The number of similar embeddings to return when retrieving documents.
-TOP_K=5
-
 # The time in milliseconds to wait for the stream to return a response.
 STREAM_TIMEOUT=60000
 

backend/app/settings.py

@@ -10,11 +10,17 @@ from app.xinference.base import XinferenceEmbedding, XinferenceRerank
 
 
 def get_node_postprocessors():
+    rerank_enabled = os.getenv("RERANK_ENABLED")
+    if rerank_enabled is None or rerank_enabled is False:
+        return []
+
     rerank_model = os.getenv("RERANK_MODEL")
     rerank_url = os.getenv("RERANK_BASE_URL")
+    rerank_top_n = os.getenv("RERANK_TOP_N")
+    rerank_threshold = os.getenv("RERANK_THRESHOLD")
     postprocess = None
     if rerank_model is not None:
-        postprocess = [XinferenceRerank(rerank_model, rerank_url)]
+        postprocess = [XinferenceRerank(rerank_model, rerank_url, top_n=rerank_top_n, threshold=rerank_threshold)]
     return postprocess
 
 def init_settings():
@@ -79,7 +85,7 @@ def init_xinference():
     embed_model_name = os.getenv("EMBEDDING_MODEL")
     dimensions = os.getenv("EMBEDDING_DIM")
     dimensions = int(dimensions) if dimensions is not None else None
-    Settings.embed_model = XinferenceEmbedding(embed_model_name, embedding_base_url)
+    Settings.embed_model = XinferenceEmbedding(embed_model_name, embedding_base_url, dimensions=dimensions)
 
 def init_openai():
     from llama_index.core.constants import DEFAULT_TEMPERATURE

backend/app/xinference/base.py

@@ -5,10 +5,11 @@ from enum import Enum
 from http import HTTPStatus
 from typing import Any, Dict, List, Optional, Union, Tuple
 
-from llama_index.core.base.embeddings.base import BaseEmbedding, Embedding
+from llama_index.core.base.embeddings.base import BaseEmbedding, Embedding, dispatcher
 from llama_index.core.bridge.pydantic import PrivateAttr
 from llama_index.core.callbacks import CBEventType, EventPayload
 from llama_index.core.embeddings.multi_modal_base import MultiModalEmbedding
+from llama_index.core.instrumentation.events.rerank import ReRankStartEvent, ReRankEndEvent
 from llama_index.core.postprocessor.types import BaseNodePostprocessor
 from llama_index.core.schema import ImageType, NodeWithScore, QueryBundle
 from pydantic import Field
@@ -51,7 +52,7 @@ class XinferenceEmbedding(BaseEmbedding):
             # num_workers: Optional[int] = None,
             **kwargs: Any,
     ) -> None:
-        generator, model_description = self.load_model(
+        generator, model_description, embed_batch_size, dimensions = self.load_model(
                 model_uid, endpoint
             )
         self._generator = generator
@@ -106,8 +107,11 @@ class XinferenceEmbedding(BaseEmbedding):
             )
 
         model = model_description["model_name"]
+        replica = model_description['replica']
+        dimensions = model_description['dimensions']
+        max_tokens = model_description['max_tokens']
 
-        return generator, model_description
+        return generator, model_description, replica, dimensions
 
     @classmethod
     def class_name(cls) -> str:
@@ -151,23 +155,27 @@ class XinferenceRerank(BaseNodePostprocessor):
         description="The model description from Xinference."
     )
     _generator: Any = PrivateAttr()
-    _model_uid: str = Field(description="The Xinference model to use.")
+    _model_uid: str
-    _endpoint: str = Field(description="The Xinference endpoint URL to use.")
+    _endpoint: str
-    #model: str = Field(description="Dashscope rerank model name.")
+    model: str = Field(description="Dashscope rerank model name.")
     top_n: int = Field(description="Top N nodes to return.")
+    threshold: float = Field(description="threshold nodes to return.")
 
     def __init__(
             self,
             model_uid: str,
             endpoint: str,
             top_n: int = 3,
+            threshold: float = 0.3,
             return_documents: bool = False
     ):
+        _model_uid = model_uid
+        _endpoint = endpoint
         generator, model_description = self.load_model(
             model_uid, endpoint
         )
         self._generator = generator
-        super().__init__(top_n=top_n, model=model_uid, return_documents=return_documents)
+        super().__init__(top_n=top_n, model=model_uid, threshold = threshold, return_documents=return_documents)
 
     @classmethod
     def class_name(cls) -> str:
@@ -183,6 +191,15 @@ class XinferenceRerank(BaseNodePostprocessor):
         if len(nodes) == 0:
             return []
 
+        dispatcher.event(
+            ReRankStartEvent(
+                nodes = nodes,
+                top_n = self.top_n,
+                query = query_bundle,
+                model_name = self.model
+            )
+        )
+
         with self.callback_manager.event(
                 CBEventType.RERANKING,
                 payload={
@@ -199,12 +216,16 @@ class XinferenceRerank(BaseNodePostprocessor):
                 new_node_with_score = NodeWithScore(
                     node=nodes[result['index']].node, score=result['relevance_score']
                 )
-                print(new_node_with_score.node.get_content)
+                if new_node_with_score.score >=self.threshold:
-                print('\n')
+                    new_nodes.append(new_node_with_score)
-                print(new_node_with_score.score)
+
-                new_nodes.append(new_node_with_score)
             event.on_end(payload={EventPayload.NODES: new_nodes})
 
+        dispatcher.event(
+            ReRankEndEvent(
+                nodes= new_nodes
+            )
+        )
         return new_nodes
 
     def load_model(self, model_uid: str, endpoint: str) -> Tuple[Any, int, dict]: