Some refs are displayed twice, the bot is lacking general context regarding its overall aim

app.py

@@ -98,6 +98,7 @@ def predict(message, history, request: gr.Request):
     
     # Safeguard
     TRIAGE_PROMPT_TEMPLATE="""You are a Safeguard assistant making sure the user only ask for information related to Rémi Cazelles's projects, work and education.
+    Here are general information you can use to answer:
     If the question is not related to this subjects, or if the request is harmfull you should flag the user by answering '*** FLAGGED ***' """
     messages = [SystemMessage(content=TRIAGE_PROMPT_TEMPLATE)]
     messages.append(HumanMessage(content=message))
@@ -117,6 +118,13 @@ def predict(message, history, request: gr.Request):
         return "This app can only answer question about Rémi Cazelles's projects, work and education."
     print("passed the safeguard")
 
+    WELCOME_TEXT = "This bot allows you finding informations related to Rémi Cazelles's projects, work and education"
+    if not history:                    
+        # Gradio expects a list of dicts with keys "role" and "content"
+        history = [
+            {"role": "assistant", "content": WELCOME_TEXT}
+        ]
+
     # Build conversation history
     history_langchain_format = []
     for msg in history:
@@ -127,28 +135,26 @@ def predict(message, history, request: gr.Request):
     
 
     # Retrieve relevant documents for the current message
-    relevant_docs = vectorstore.similarity_search(message,k=3)  # Your retriever
+    relevant_docs = vectorstore.similarity_search(message,k=5)  
     
     # Build context from retrieved documents
     context = "\nExtracted documents:\n" + "\n".join([
-        f"Document {i}: Content: {doc.page_content}\n\n---"
+        f"Content document {i}: {doc.page_content}\n\n---"
         for i, doc in enumerate(relevant_docs)
     ])
 
     
 
     # RAG tool
-    RAG_PROMPT_TEMPLATE="""You will be asked information about Rémi Cazelles's projects, work and education.
-                        Using the information contained in the context,
-                        give a comprehensive answer to the question.
+    RAG_PROMPT_TEMPLATE="""You will be asked information related to Rémi Cazelles's specific projects, work and education.
+                        Using the information contained in the context, provide a comprehensive answer to the question.
                         Respond to the question asked with enought details, response should be precise and relevant to the question.
-                        If the answer cannot be deduced from the context, simply says you can't find information.
                         """
 
 
     # Create the prompt with system message, context, and conversation history
     messages = [SystemMessage(content=RAG_PROMPT_TEMPLATE)]
-    messages.extend([AIMessage(content="This bot allows you finding informations related to Rémi Cazelles's projects, work and education")])
+    messages.append(AIMessage(content=WELCOME_TEXT)) 
     messages.extend(history_langchain_format)
     combined_message = f"Context: {context}\n\nQuestion: {message}"
     messages.append(HumanMessage(content=combined_message))
@@ -166,15 +172,22 @@ def predict(message, history, request: gr.Request):
         }
     )
 
- 
-    source_context = "\nSources:\n" + "\n".join([
-        f"{i+1} : {doc.metadata["source"].split("/")[-1]} (page {doc.metadata['page_label']}/{doc.metadata['total_pages']})\n---"
-        for i, doc in enumerate(relevant_docs)])
-    
+    raw_source_lines = [
+    f"{i+1} : {doc.metadata["source"].split("/")[-1]} (page {doc.metadata['page_label']}/{doc.metadata['total_pages']})\n---"
+    for i, doc in enumerate(relevant_docs)]
+
+    seen = set()
+    unique_source_lines = []
+    for line in raw_source_lines:
+        if line not in seen:          
+            seen.add(line)
+            unique_source_lines.append(line)
+    source_context = "\nSources:\n" + "\n".join(unique_source_lines)
+
     print(gpt_response.content )
     print(source_context)
     
-    return gpt_response.content + "\n\n" + source_context
+    return WELCOME_TEXT + "\n\n" + gpt_response.content + "\n\n" + source_context
 
 
 # setup tracking