Added RAG settings to settings.py, vector_store and chat_service to add similarity_top_k and similarity_score

private_gpt/components/vector_store/vector_store_component.py

@@ -135,6 +135,7 @@ class VectorStoreComponent:
         similarity_top_k: int = 2,
     ) -> VectorIndexRetriever:
         # This way we support qdrant (using doc_ids) and the rest (using filters)
+        similarity_top_k = self.settings.rag.similarity_top_k
         return VectorIndexRetriever(
             index=index,
             similarity_top_k=similarity_top_k,

private_gpt/server/chat/chat_service.py

@@ -8,6 +8,9 @@ from llama_index.core.chat_engine.types import (
 from llama_index.core.indices import VectorStoreIndex
 from llama_index.core.indices.postprocessor import MetadataReplacementPostProcessor
 from llama_index.core.llms import ChatMessage, MessageRole
+from llama_index.core.postprocessor import (
+    SimilarityPostprocessor,
+)
 from llama_index.core.storage import StorageContext
 from llama_index.core.types import TokenGen
 from pydantic import BaseModel
@@ -20,6 +23,7 @@ from private_gpt.components.vector_store.vector_store_component import (
 )
 from private_gpt.open_ai.extensions.context_filter import ContextFilter
 from private_gpt.server.chunks.chunks_service import Chunk
+from private_gpt.settings.settings import Settings
 
 
 class Completion(BaseModel):
@@ -68,14 +72,18 @@ class ChatEngineInput:
 
 @singleton
 class ChatService:
+    settings: Settings
+
     @inject
     def __init__(
         self,
+        settings: Settings,
         llm_component: LLMComponent,
         vector_store_component: VectorStoreComponent,
         embedding_component: EmbeddingComponent,
         node_store_component: NodeStoreComponent,
     ) -> None:
+        self.settings = settings
         self.llm_component = llm_component
         self.embedding_component = embedding_component
         self.vector_store_component = vector_store_component
@@ -98,6 +106,7 @@ class ChatService:
         use_context: bool = False,
         context_filter: ContextFilter | None = None,
     ) -> BaseChatEngine:
+        settings = self.settings
         if use_context:
             vector_index_retriever = self.vector_store_component.get_retriever(
                 index=self.index, context_filter=context_filter
@@ -108,6 +117,9 @@ class ChatService:
                 llm=self.llm_component.llm,  # Takes no effect at the moment
                 node_postprocessors=[
                     MetadataReplacementPostProcessor(target_metadata_key="window"),
+                    SimilarityPostprocessor(
+                        similarity_cutoff=settings.rag.similarity_value
+                    ),
                 ],
             )
         else:

private_gpt/settings/settings.py

@@ -280,6 +280,17 @@ class UISettings(BaseModel):
     )
 
 
+class RagSettings(BaseModel):
+    similarity_top_k: int = Field(
+        2,
+        description="This value controls the number of documents returned by the RAG pipeline",
+    )
+    similarity_value: float = Field(
+        None,
+        description="If set, any documents retrieved from the RAG must meet a certain match score. Acceptable values are between 0 and 1.",
+    )
+
+
 class PostgresSettings(BaseModel):
     host: str = Field(
         "localhost",
@@ -375,6 +386,7 @@ class Settings(BaseModel):
     azopenai: AzureOpenAISettings
     vectorstore: VectorstoreSettings
     nodestore: NodeStoreSettings
+    rag: RagSettings
     qdrant: QdrantSettings | None = None
     postgres: PostgresSettings | None = None
 

settings.yaml

@@ -42,6 +42,12 @@ llm:
   tokenizer: mistralai/Mistral-7B-Instruct-v0.2
   temperature: 0.1      # The temperature of the model. Increasing the temperature will make the model answer more creatively. A value of 0.1 would be more factual. (Default: 0.1)
 
+rag:
+  similarity_top_k: 2
+  #This value controls how many "top" documents the RAG returns to use in the context.
+  #similarity_value: 0.45
+  #This value is disabled by default.  If you enable this settings, the RAG will only use articles that meet a certain percentage score.
+
 llamacpp:
   prompt_style: "mistral"
   llm_hf_repo_id: TheBloke/Mistral-7B-Instruct-v0.2-GGUF