Rename tools/orchestrator to tools/orchestrator.py

tools/orchestrator.py

@@ -0,0 +1,106 @@
+import os
+import json
+import re
+import logging
+import time
+
+from model_logic import call_model_stream, MODELS_BY_PROVIDER, get_default_model_display_name_for_provider
+from memory_logic import retrieve_memories_semantic, retrieve_rules_semantic
+from tools.websearch import search_and_scrape_duckduckgo, scrape_url
+import prompts
+from utils import format_insights_for_prompt
+
+logger = logging.getLogger(__name__)
+
+WEB_SEARCH_ENABLED = os.getenv("WEB_SEARCH_ENABLED", "true").lower() == "true"
+TOOL_DECISION_PROVIDER = os.getenv("TOOL_DECISION_PROVIDER", "groq")
+TOOL_DECISION_MODEL_ID = os.getenv("TOOL_DECISION_MODEL", "llama3-8b-8192")
+MAX_HISTORY_TURNS = int(os.getenv("MAX_HISTORY_TURNS", 7))
+
+def decide_on_tool(user_input: str, chat_history_for_prompt: list, initial_insights_ctx_str: str):
+    user_input_lower = user_input.lower()
+
+    if "http://" in user_input or "https://" in user_input:
+        url_match = re.search(r'(https?://[^\s]+)', user_input)
+        if url_match: return "scrape_url_and_report", {"url": url_match.group(1)}
+
+    if len(user_input.split()) <= 3 and any(kw in user_input_lower for kw in ["hello", "hi", "thanks", "ok", "bye"]) and not "?" in user_input:
+        return "quick_respond", {}
+    
+    if len(user_input.split()) > 3 or "?" in user_input or any(w in user_input_lower for w in ["what is", "how to", "explain", "search for"]):
+        history_snippet = "\n".join([f"{msg['role']}: {msg['content'][:100]}" for msg in chat_history_for_prompt[-2:]])
+        guideline_snippet = initial_insights_ctx_str[:200].replace('\n', ' ')
+        tool_user_prompt = prompts.get_tool_user_prompt(user_input, history_snippet, guideline_snippet)
+        tool_decision_messages = [{"role":"system", "content": prompts.TOOL_SYSTEM_PROMPT}, {"role":"user", "content": tool_user_prompt}]
+        tool_model_display = next((dn for dn, mid in MODELS_BY_PROVIDER.get(TOOL_DECISION_PROVIDER.lower(), {}).get("models", {}).items() if mid == TOOL_DECISION_MODEL_ID), None)
+        if not tool_model_display: tool_model_display = get_default_model_display_name_for_provider(TOOL_DECISION_PROVIDER)
+        
+        if tool_model_display:
+            try:
+                tool_resp_raw = "".join(list(call_model_stream(provider=TOOL_DECISION_PROVIDER, model_display_name=tool_model_display, messages=tool_decision_messages, temperature=0.0, max_tokens=150)))
+                json_match_tool = re.search(r"\{.*\}", tool_resp_raw, re.DOTALL)
+                if json_match_tool:
+                    action_data = json.loads(json_match_tool.group(0))
+                    action_type = action_data.get("action", "quick_respond")
+                    action_input = action_data.get("action_input", {})
+                    if not isinstance(action_input, dict): action_input = {}
+                    return action_type, action_input
+            except Exception as e:
+                logger.error(f"Tool decision LLM error: {e}")
+
+    return "quick_respond", {}
+
+def orchestrate_and_respond(user_input: str, provider_name: str, model_display_name: str, chat_history_for_prompt: list[dict], custom_system_prompt: str = None, ui_api_key_override: str = None):
+    process_start_time = time.time()
+    request_id = os.urandom(4).hex()
+    logger.info(f"ORCHESTRATOR [{request_id}] Start. User: '{user_input[:50]}...'")
+
+    history_str_for_prompt = "\n".join([f"{('User' if t['role'] == 'user' else 'AI')}: {t['content']}" for t in chat_history_for_prompt[-(MAX_HISTORY_TURNS * 2):]])
+    
+    yield "status", "[Checking guidelines...]"
+    initial_insights = retrieve_rules_semantic(f"{user_input}\n{history_str_for_prompt}", k=5)
+    initial_insights_ctx_str, parsed_initial_insights_list = format_insights_for_prompt(initial_insights)
+    
+    yield "status", "[Choosing best approach...]"
+    action_type, action_input_dict = decide_on_tool(user_input, chat_history_for_prompt, initial_insights_ctx_str)
+    logger.info(f"ORCHESTRATOR [{request_id}]: Tool Decision: Action='{action_type}', Input='{action_input_dict}'")
+    
+    yield "status", f"[Path: {action_type}]"
+    final_system_prompt_str = custom_system_prompt or prompts.DEFAULT_SYSTEM_PROMPT
+    context_str, final_user_prompt_str = None, ""
+
+    if action_type == "answer_using_conversation_memory":
+        yield "status", "[Searching conversation memory...]"
+        mems = retrieve_memories_semantic(f"User query: {user_input}\nContext:\n{history_str_for_prompt[-1000:]}", k=2)
+        context_str = "Relevant Past Interactions:\n" + "\n".join([f"- User:{m.get('user_input','')}->AI:{m.get('bot_response','')} (Takeaway:{m.get('metrics',{}).get('takeaway','N/A')})" for m in mems]) if mems else "No relevant past interactions found."
+        final_system_prompt_str += " Respond using Memory Context, guidelines, & history."
+    elif WEB_SEARCH_ENABLED and action_type in ["search_duckduckgo_and_report", "scrape_url_and_report"]:
+        query_or_url = action_input_dict.get("search_engine_query") or action_input_dict.get("url")
+        if query_or_url:
+            yield "status", f"[Web: '{query_or_url[:60]}'...]"
+            web_results = []
+            try:
+                if action_type == "search_duckduckgo_and_report": web_results = search_and_scrape_duckduckgo(query_or_url, num_results=2)
+                elif action_type == "scrape_url_and_report": web_results = [scrape_url(query_or_url)]
+            except Exception as e: web_results = [{"url": query_or_url, "error": str(e)}]
+            context_str = "Web Content:\n" + "\n".join([f"Source {i+1}:\nURL:{r.get('url','N/A')}\nTitle:{r.get('title','N/A')}\nContent:\n{(r.get('content') or r.get('error') or 'N/A')[:3500]}\n---" for i,r in enumerate(web_results)]) if web_results else f"No results from {action_type} for '{query_or_url}'."
+            yield "status", "[Synthesizing web report...]"
+            final_system_prompt_str += " Generate report/answer from web content, history, & guidelines. Cite URLs as [Source X]."
+    else: # quick_respond or fallback
+        final_system_prompt_str += " Respond directly using guidelines & history."
+
+    final_user_prompt_str = prompts.get_final_response_prompt(history_str_for_prompt, initial_insights_ctx_str, user_input, context_str)
+    final_llm_messages = [{"role": "system", "content": final_system_prompt_str}, {"role": "user", "content": final_user_prompt_str}]
+    
+    streamed_response = ""
+    try:
+        for chunk in call_model_stream(provider=provider_name, model_display_name=model_display_name, messages=final_llm_messages, api_key_override=ui_api_key_override, temperature=0.6, max_tokens=2500):
+            if isinstance(chunk, str) and chunk.startswith("Error:"):
+                streamed_response += f"\n{chunk}\n"; yield "response_chunk", f"\n{chunk}\n"; break
+            streamed_response += chunk; yield "response_chunk", chunk
+    except Exception as e:
+        streamed_response += f"\n\n(Error: {e})"; yield "response_chunk", f"\n\n(Error: {e})"
+    
+    final_bot_text = streamed_response.strip() or "(No response or error.)"
+    logger.info(f"ORCHESTRATOR [{request_id}]: Finished. Total: {time.time() - process_start_time:.2f}s. Resp len: {len(final_bot_text)}")
+    yield "final_response_and_insights", {"response": final_bot_text, "insights_used": parsed_initial_insights_list}