Improved no-context response and logs

app.py

@@ -72,18 +72,11 @@ def completion(history, model, system_prompt: str, tools=None, chat_params=chat_
     }
     if tools:
         request_params.update({"tool_choice": "auto", "tools": tools})
-    
-    cprint("=" * 150, "green")
-    print(json.dumps(request_params, indent=2, ensure_ascii=False))
-    cprint("=" * 150, "green")
 
     return client.chat.completions.create(**request_params)  
 
 
 def llm_in_loop(history, system_prompt, recursive):  
-    
-    cprint(history, "cyan")
-    cprint("~" * 150, "yellow")
 
     try:   
         models = client.models.list()
@@ -113,6 +106,7 @@ def llm_in_loop(history, system_prompt, recursive):
             history[-1].content += chunk.choices[0].delta.content
             yield history[recursive:]
     
+    # Convert arguments to a valid JSON
     arguments = clean_json_string(arguments) if arguments else "{}"
     arguments = json.loads(arguments)
 
@@ -122,7 +116,6 @@ def llm_in_loop(history, system_prompt, recursive):
     if name:
         try:
             result = str(tools[name].invoke(input=arguments))
-            cprint(f"*** Tool {name} invoked with arguments {arguments}, result: {result}", "yellow")
 
         except Exception as err:
             result = f"💥 Error: {err}"
@@ -137,6 +130,7 @@ def llm_in_loop(history, system_prompt, recursive):
 
 
 def respond(message, history, additional_inputs):  
+    
     history.append(ChatMessage(role="user", content=message))
     yield from llm_in_loop(history, additional_inputs, -1)
 
@@ -144,6 +138,7 @@ def respond(message, history, additional_inputs):
 
 if __name__ == "__main__":
 
-    system_prompt = gr.Textbox(label="System prompt", value=cfg.system_prompt_template, lines=10)
+    # system_prompt = gr.State(value=cfg.system_prompt_template)
+    system_prompt = gr.Textbox(label="System prompt", value=cfg.system_prompt_template, lines=10, visible=False)
     demo = gr.ChatInterface(respond, type="messages", additional_inputs=[system_prompt])
     demo.launch()

config.yaml

@@ -5,14 +5,22 @@ vdb:
   embeddings_model: BAAI/bge-m3
   number_of_contexts: 4
   vs_local_path: data/vdb
-  embedding_score_threshold: 0.4
+  embedding_score_threshold: 0.3
 
   # Context formatting parameters
   context_fmt: "Context document {num_document}:\n{document_content}"
   join_str: "\n\n"
   header_context_str: "The following is the context to help you answer the question (sorted from most to least relevant):\n\n"
   footer_context_str: "\n\nAnswer based only on the above context."
-  no_context_str: "Answer 'no relevant context found'."
+  no_context_str: | 
+    Answer exactly with the following text respecting HTML tags: 
+    "No relevant context found. Here are Aina Kit and Discord links for more information:
+    1. Aina Kit official: <a href="https://langtech-bsc.gitbook.io/aina-kit">https://langtech-bsc.gitbook.io/aina-kit</a>
+    2. Discord community: <a href="https://discord.com/invite/twy3Gn">https://discord.com/invite/twy3Gn</a>"
+
+  # https://discord.com/invite/twy3GnBCaY
+  # https://discord.com/invite/twy3GnBCaY
+  # https://langtech-bsc.gitbook.io/aina-kit
 
 # LLM client configuration
 # ================================================================================

src/tools.py

@@ -4,6 +4,8 @@ from typing import Dict, Union, get_origin, get_args
 from pydantic import BaseModel, Field
 from types import UnionType
 import logging
+log = logging.getLogger(__name__)
+
 from src.vectorstore import VectorStore
 from omegaconf import OmegaConf
 
@@ -104,14 +106,14 @@ def tool_register(cls: BaseModel):
 @tool_register
 class retrieve_aina_data(ToolBase):
     """Retrieves relevant information from Aina Challenge vectorstore, based on the user's query."""
-    logging.info("@tool_register: retrieve_aina_data()")
+    log.info("@tool_register: retrieve_aina_data()")
 
     query: str = Field(description="The user's input or question, used to search in Aina Challenge vectorstore.")
-    logging.info(f"query: {query}")
+    log.info(f"query: {query}")
 
     @classmethod
     def invoke(cls, input: Dict) -> str:
-        logging.info(f"retrieve_aina_data.invoke() input: {input}")
+        log.info(f"retrieve_aina_data.invoke() input: {input}")
         
         # Check if the input is a dictionary
         query = input.get("query", None)

src/vectorstore.py

@@ -3,6 +3,7 @@ from langchain_community.vectorstores import FAISS
 from langchain_huggingface import HuggingFaceEmbeddings
 from huggingface_hub import snapshot_download
 import logging
+log = logging.getLogger(__name__)
 
 from termcolor import cprint
 
@@ -55,7 +56,7 @@ class VectorStore:
             Defaults to "\n\nAnswer based only on the above context.".
         """
         
-        logging.info("Loading vectorstore...")
+        log.info("Loading vectorstore...")
 
         # Retrieval parameters
         self.number_of_contexts = number_of_contexts
@@ -69,22 +70,22 @@ class VectorStore:
         self.no_context_str = no_context_str
 
         embeddings = HuggingFaceEmbeddings(model_name=embeddings_model)
-        logging.info(f"Loaded embeddings model: {embeddings_model}")
+        log.info(f"Loaded embeddings model: {embeddings_model}")
 
         if vs_hf_path:
             hf_vectorstore = snapshot_download(repo_id=vs_hf_path)
             self.vdb = FAISS.load_local(hf_vectorstore, embeddings, allow_dangerous_deserialization=True)
-            logging.info(f"Loaded vectorstore from {vs_hf_path}")
+            log.info(f"Loaded vectorstore from {vs_hf_path}")
         else:
             self.vdb = FAISS.load_local(vs_local_path, embeddings, allow_dangerous_deserialization=True)
-            logging.info(f"Loaded vectorstore from {vs_local_path}")
+            log.info(f"Loaded vectorstore from {vs_local_path}")
 
     
     def get_context(self, query,):
 
         # Retrieve documents
         results = self.vdb.similarity_search_with_relevance_scores(query=query, k=self.number_of_contexts, score_threshold=self.embedding_score_threshold)
-        logging.info(f"Retrieved {len(results)} documents from the vectorstore.")
+        log.info(f"Retrieved {len(results)} documents from the vectorstore.")
 
         # Return formatted context
         return self._beautiful_context(results)
@@ -92,7 +93,7 @@ class VectorStore:
     
     def _beautiful_context(self, docs):
         
-        logging.info(f"Formatting {len(docs)} contexts...")
+        log.info(f"Formatting {len(docs)} contexts...")
 
         # If no documents are retrieved, return the no_context_str
         if not docs:
@@ -101,6 +102,8 @@ class VectorStore:
         contexts = []
         for i, doc in enumerate(docs):
             
+            log.info(f"Document {i+1} (score: {doc[1]:.4f}): {repr(doc[0].page_content[:100])}...")
+            
             # Format each context document using the provided template
             context = self.context_fmt.format(num_document=i + 1, document_content=doc[0].page_content)
             contexts.append(context)