Remove prompt instruction tags & reconfigure streaming newlines

memri_bot/plugin.py

@@ -111,7 +111,7 @@ class MemriBot(PluginBase):
             return Response(status_code=400, content="rss_entry_id is missing")
 
         # append instructions
-        context += f"\n\n<instructions>\n{chat.system}"
+        context += f"\n\n{chat.system}"
         # truncate final query if it is longer that QUERY_MAX_TOKENS
         chat.messages[-1].content = truncate_with_textrank(chat.messages[-1].content, QUERY_MAX_TOKENS)
         messages = [{"role": str(m.role), "content": m.content} for m in chat.messages]
@@ -171,7 +171,7 @@ class MemriBot(PluginBase):
         query = truncate_with_textrank(query, QUERY_MAX_TOKENS)
         context = truncate_with_textrank(context, CONTEXT_MAX_TOKENS)
         # append instructions
-        context += f"\n\n<instructions>\n{system}"
+        context += f"\n\n{system}"
         logger.info(f"Sending LLM request '{query}' with context '{context}'")
         return self.get_llm_response(query, context, stream, session)
 
@@ -224,7 +224,7 @@ class MemriBot(PluginBase):
 
         response_text = ""
         for line in self.streaming_request(context, messages=[m.to_dict() for m in messages]):
-            stream_chunk = json.loads(line)
+            stream_chunk = json.loads(line.rstrip('\n'))
             text = stream_chunk["choices"][0]["text"]
             response_text += text
             yield line + '\n'
@@ -252,5 +252,4 @@ class MemriBot(PluginBase):
         with s.post(f"{LLM_API_URL}/v1/chat/completions", stream=True, json=data, headers=headers) as r:
             for line in r.iter_lines():
                 line = line.decode("utf-8")
-                line = line.strip()
-                yield line
\ No newline at end of file
+                yield line + '\n'
\ No newline at end of file