Update main.py

main.py

@@ -15,25 +15,43 @@ SEARCH_API_URL = "https://rkihacker-brave.hf.space/search"
 MODEL_NAME = "Binglity-Lite"
 BACKEND_MODEL = "meta-llama/llama-3.1-8b-instruct/fp-8"
 
-# --- A More Advanced System Prompt ---
+# --- Final Advanced System Prompt ---
 SYSTEM_PROMPT = """
-You are "Binglity-Lite", a state-of-the-art AI assistant. Your purpose is to provide accurate, unbiased, and comprehensive answers by synthesizing information from real-time web search results.
-
-**Your Instructions:**
-1.  **Analyze the User's Query**: Deeply understand the user's question, intent, and the specific information they are seeking.
-2.  **Synthesize, Don't List**: Do not simply list or summarize the search results. Your primary task is to integrate the information from the multiple sources provided into a single, cohesive, and well-structured response.
-3.  **Be Factual and Unbiased**: Base your entire response ONLY on the information contained within the provided search results. Do not introduce any external knowledge or personal opinions.
-4.  **Handle Contradictions**: If the search results present conflicting information, acknowledge the discrepancy and present the different viewpoints as found in the sources.
-5.  **Address Insufficient Information**: If the search results do not contain enough information to provide a complete answer, explicitly state that. Do not speculate or fill in the gaps.
-6.  **Maintain a Helpful Tone**: Your persona is knowledgeable, helpful, and neutral.
-7.  **Structure for Clarity**: Use clear language and logical formatting (like paragraphs or bullet points if appropriate) to make the information easy to understand.
+You are "Binglity-Lite", a highly advanced AI search assistant. Your purpose is to provide users with accurate, comprehensive, and trustworthy answers by synthesizing information from a given set of web search results.
+
+**Core Directives:**
+
+1.  **Answer Directly**: Immediately address the user's question. **Do not** use introductory phrases like "Based on the search results..." or "Here is the information I found...". Your tone should be confident, objective, and encyclopedic.
+
+2.  **Synthesize, Don't Summarize**: Your primary task is to weave information from multiple sources into a single, cohesive, and well-structured answer. Do not simply describe what each source says one by one.
+
+3.  **Cite with Inline Markdown Links**: This is your most important instruction. When you present a fact or a piece of information from a source, you **must** cite it immediately using an inline Markdown link.
+    *   **Format**: The format must be `[phrase or sentence containing the fact](URL)`. The URL must come from the `URL:` field of the provided source.
+    *   **Example**: If a source with URL `https://example.com/science` says "The Earth is the third planet from the Sun", your output should be: "The Earth is the [third planet from the Sun](https://example.com/science)."
+    *   **Rule**: Every piece of information in your answer must be attributable to a source via these inline links.
+
+4.  **Be Fact-Based**: Your entire response must be based **exclusively** on the information provided in the search results. Do not use any outside knowledge.
+
+5.  **Filter for Relevance**: If a search result is not relevant to the user's query, ignore it completely. Do not mention it in your response.
+
+6.  **Handle Ambiguity**: If the search results are contradictory or insufficient to answer the question fully, state this clearly in your response, citing the conflicting sources.
+
+**Final Output Structure:**
+
+Your final response MUST be structured in two parts:
+
+1.  **The Synthesized Answer**: A well-written response that directly answers the user's query, with facts and statements properly cited using inline Markdown links as described above.
+
+2.  **Sources Section**: After the answer, add a section header `## Sources`. Under this header, provide a bulleted list of the full titles and URLs of every source you used.
+    *   **Format**: `- [Title of Source](URL)`
 """
 
+
 # --- FastAPI App ---
 app = FastAPI(
     title="Binglity-Lite API",
     description="A web search-powered, streaming-capable chat completions API.",
-    version="1.1.0",
+    version="1.2.0",
 )
 
 # --- Pydantic Models for OpenAI Compatibility ---
@@ -67,7 +85,7 @@ async def perform_web_search(query: str) -> List[Dict[str, Any]]:
 
 def format_search_results_for_prompt(results: List[Dict[str, Any]]) -> str:
     if not results:
-        return "No relevant search results were found. Please inform the user that you could not find information on their query."
+        return "No relevant search results were found. Inform the user that you could not find information on their query."
     
     formatted = "### Web Search Results ###\n\n"
     for i, result in enumerate(results):
@@ -79,16 +97,11 @@ def format_search_results_for_prompt(results: List[Dict[str, Any]]) -> str:
 
 # --- Streaming Logic ---
 async def stream_response_generator(payload: Dict[str, Any]) -> AsyncGenerator[str, None]:
-    """
-    Yields chunks from the inference API, formatted for OpenAI compatibility.
-    """
     headers = {
         "Authorization": f"Bearer {INFERENCE_API_KEY}",
         "Content-Type": "application/json",
         "Accept": "text/event-stream"
     }
-    
-    # Create a unique ID for the response stream
     response_id = f"chatcmpl-{uuid.uuid4()}"
     created_time = int(time.time())
 
@@ -98,33 +111,22 @@ async def stream_response_generator(payload: Dict[str, Any]) -> AsyncGenerator[s
                 error_content = await response.aread()
                 raise HTTPException(status_code=response.status_code, detail=f"Error from inference API: {error_content.decode()}")
 
-            # Stream the response line by line
             async for line in response.aiter_lines():
                 if line.startswith("data:"):
                     line_data = line[5:].strip()
                     if line_data == "[DONE]":
-                        # Send the final data chunk and the done message
                         yield f"data: {json.dumps({'id': response_id, 'model': MODEL_NAME, 'object': 'chat.completion.chunk', 'created': created_time, 'choices': [{'index': 0, 'delta': {}, 'finish_reason': 'stop'}]})}\n\n"
                         yield "data: [DONE]\n\n"
                         break
                     
                     try:
                         chunk = json.loads(line_data)
-                        # Reformat the chunk to be OpenAI compliant
                         formatted_chunk = {
-                            "id": response_id,
-                            "object": "chat.completion.chunk",
-                            "created": created_time,
-                            "model": MODEL_NAME,
-                            "choices": [{
-                                "index": 0,
-                                "delta": chunk["choices"][0].get("delta", {}),
-                                "finish_reason": chunk["choices"][0].get("finish_reason")
-                            }]
+                            "id": response_id, "object": "chat.completion.chunk", "created": created_time, "model": MODEL_NAME,
+                            "choices": [{"index": 0, "delta": chunk["choices"][0].get("delta", {}), "finish_reason": chunk["choices"][0].get("finish_reason")}]
                         }
                         yield f"data: {json.dumps(formatted_chunk)}\n\n"
                     except json.JSONDecodeError:
-                        print(f"Could not decode JSON from line: {line_data}")
                         continue
 
 # --- API Endpoint ---
@@ -137,51 +139,32 @@ async def chat_completions(request: ChatCompletionRequest):
     if not user_query or request.messages[-1].role.lower() != 'user':
         raise HTTPException(status_code=400, detail="The last message must be from the 'user' and contain content.")
 
-    # 1. Perform Web Search
     search_results = await perform_web_search(user_query)
     formatted_results = format_search_results_for_prompt(search_results)
     
-    # 2. Construct prompt for the backend model
-    final_user_prompt = f"User's question: \"{user_query}\"\n\nBased ONLY on the provided search results below, answer the user's question.\n\n{formatted_results}"
+    final_user_prompt = f"User's question: \"{user_query}\"\n\nUse the web search results below to answer the user's question. Follow all rules in your system prompt exactly.\n\n{formatted_results}"
     
-    # 3. Prepare payload for Inference API
     payload = {
         "model": BACKEND_MODEL,
         "messages": [
             {"role": "system", "content": SYSTEM_PROMPT},
             {"role": "user", "content": final_user_prompt},
         ],
-        "max_tokens": request.max_tokens,
-        "temperature": request.temperature,
-        "stream": request.stream,
+        "max_tokens": request.max_tokens, "temperature": request.temperature, "stream": request.stream,
     }
 
-    # 4. Handle streaming or single response
     if request.stream:
         return StreamingResponse(stream_response_generator(payload), media_type="text/event-stream")
     else:
-        # Standard non-streaming request
         headers = {"Authorization": f"Bearer {INFERENCE_API_KEY}"}
         async with httpx.AsyncClient(timeout=120.0) as client:
             try:
                 response = await client.post(INFERENCE_API_URL, json=payload, headers=headers)
                 response.raise_for_status()
                 model_response = response.json()
-
-                # Format response to be OpenAI API compliant
                 return {
-                    "id": model_response.get("id", f"chatcmpl-{uuid.uuid4()}"),
-                    "object": "chat.completion",
-                    "created": model_response.get("created", int(time.time())),
-                    "model": MODEL_NAME,
-                    "choices": [{
-                        "index": 0,
-                        "message": {
-                            "role": "assistant",
-                            "content": model_response["choices"][0]["message"]["content"],
-                        },
-                        "finish_reason": "stop",
-                    }],
+                    "id": model_response.get("id", f"chatcmpl-{uuid.uuid4()}"), "object": "chat.completion", "created": model_response.get("created", int(time.time())), "model": MODEL_NAME,
+                    "choices": [{"index": 0, "message": {"role": "assistant", "content": model_response["choices"][0]["message"]["content"],}, "finish_reason": "stop",}],
                     "usage": model_response.get("usage", {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}),
                 }
             except httpx.HTTPStatusError as e: