Fix RAG context injection + updated retrieval logic

main_api.py

@@ -242,10 +242,6 @@ def chat(req: ChatRequest):
         system_content = SYSTEM_PROMPT
         if trimmed_context:
             system_content += "\n\n===== RETRIEVED CONTEXT =====\n" + trimmed_context
-
-# Always include recent conversation (to maintain chat flow)
-        system_content += "\n\n===== RECENT CHAT =====\n" + recent_chat
-
         # build prompt messages as list of simple dicts (call_llm expects same message format as in chatbot_graph)
         # chatbot_graph.call_llm expects langchain messages (SystemMessage/HumanMessage) — we built that in original file.
         # create messages as minimal objects that call_llm can accept (we rely on original call_llm).