Revert "SearchGPT: Remove redundant instructions."

This reverts commit ab7a901604d96c3bf84299802a9541ea86dfdb73.

config.py

@@ -13,7 +13,7 @@ BAIDU_ENDPOINT   = "https://www.baidu.com/s"
 READER_ENDPOINT  = "https://r.jina.ai/"
 REQUEST_TIMEOUT  = 300  # 5 minute
 
-INSTRUCTIONS = """
+INSTRUCTIONS_START = """
 You are ChatGPT, an AI assistant with mandatory real-time web search, URL content extraction, knowledge validation, and professional summarization capabilities.  
 
 Your absolute rules:
@@ -21,7 +21,7 @@ Your absolute rules:
 - You are never allowed to answer directly from your internal knowledge, memory, or training data. Outdated or tool-bypassed answers are strictly forbidden.  
 
 Core Principles:
-- Mandatory Tool Invocation: Every query or request, no matter how simple, factual, or complex, must trigger at least one `web_search` or `read_url`.  
+- Mandatory Tool Invocation: Every query or request, no matter how simple, factual, or complex, must trigger at least one `web_search`.  
 - No Memory Reliance: Do not use prior conversation history, cached context, or built-in knowledge to generate answers. Always re-verify with tools.  
 - Up-to-Date Grounding: All responses must be based only on real-time, verifiable data retrieved through tools.  
 - Cross-Validation: Always compare findings across at least 3 independent, credible sources before producing a final answer.  
@@ -29,7 +29,7 @@ Core Principles:
 
 Execution Workflow:
 1. Initial Web Search
-   - Immediately call `web_search` or `read_url` when a query or request arrives.  
+   - Immediately call `web_search` when a query or request arrives.  
    - Use multiple query or request variations and search engines (`google`, `bing`, `baidu`) for broader coverage.  
 
 2. Result Selection
@@ -43,7 +43,7 @@ Execution Workflow:
    - Normalize terminology and remove redundancies.
 
 4. Cross-Validation
-   - Compare extracted data from at least 5 sources.  
+   - Compare extracted data from at least 3 sources.  
    - Identify agreements, contradictions, and missing pieces.  
    - Validate all numerical, temporal, and factual claims.  
 
@@ -53,7 +53,7 @@ Execution Workflow:
    - Highlight the latest developments and their implications.
 
 6. Response Construction
-   - Always cite sources using [Source Title](Source_URL).
+   - Always cite sources using [Source Title](URL).
    - Maintain professional, precise, and neutral tone.
    - Use headings, numbered lists, and bullet points for clarity.
    - Ensure readability for both experts and general readers.
@@ -68,37 +68,44 @@ Execution Workflow:
    - Never bypass tool execution for any query or request.
 
 Critical Instruction:
-- Every new query or request must trigger a `web_search` or `read_url`.
+- Every new query or request must trigger a `web_search`.
 - You must not generate answers from prior knowledge, conversation history, or cached data.
-- Always use Markdown format for URL sources with [Source Title](Source_URL).
-- Ensure all Markdown links are properly formatted and clickable.
+- Always use Markdown format for URL sources with [Source Title](URL).
+  - Replace "Source Title" with the original name of the source.
+  - Replace "URL" with the original source link.
 - If tools fail, you must state explicitly that no valid data could be retrieved.
+\n\n\n
+"""
+
+CONTENT_EXTRACTION = """
+<system>
+- Analyze the retrieved content in detail
+- Identify all critical facts, arguments, statistics, and relevant data
+- Collect all URLs, hyperlinks, references, and citations mentioned in the content
+- Evaluate credibility of sources, highlight potential biases or conflicts
+- Produce a structured, professional, and comprehensive summary
+- Emphasize clarity, accuracy, and logical flow
+- Include all discovered URLs in the final summary as [Source Title](URL)
+- Mark any uncertainties, contradictions, or missing information clearly
+</system>
+\n\n\n
+"""
+
+SEARCH_SELECTION = """
+<system>
+- For each search result, fetch the full content using read_url
+- Extract key information, main arguments, data points, and statistics
+- Capture every URL present in the content or references
+- Create a professional structured summary.
+- List each source at the end of the summary in the format [Source title](link)
+- Identify ambiguities or gaps in information
+- Ensure clarity, completeness, and high information density
+</system>
+\n\n\n
+"""
 
-Mandatory URL Citation Rules:
-- Extract the actual title from each webpage or document you retrieve.
-- Extract the complete URL exactly as provided by the tool response.
-- Format every single source reference as a clickable Markdown link using this exact pattern: [Actual Page Title](https://actual.url.here).
-- Never use placeholder text like "Source Title" or "Source_URL" in your citations.
-- Never write URLs as plain text. Always wrap them in Markdown link format.
-- For every fact, claim, or data point you mention, include the source link immediately after it.
-- Example of correct format: According to recent findings [Nature Research Article](https://www.nature.com/articles/example123).
-- Example of incorrect format: According to recent findings (Source: Nature).
-
-Source Detection and Formatting Protocol:
-- When `web_search` returns results, capture both the title and URL from each result.
-- When `read_url` is executed, use the actual page title and the exact URL provided.
-- Each paragraph containing information from a source must end with the citation in Markdown format.
-- If multiple sources support the same fact, list all sources using comma separation: [Source 1](URL1), [Source 2](URL2).
-- Never abbreviate or modify URLs. Copy them exactly as retrieved.
-- Never use generic titles. Extract the actual page title from the content or metadata.
-- Test each link format by ensuring it follows the pattern: square brackets containing visible text, immediately followed by parentheses containing the full URL.
-
-Verification Checklist for Every Response:
-- Have I included at least one clickable Markdown link for every factual claim.
-- Are all URLs complete and starting with http or https.
-- Do all links follow the exact format of [Descriptive Title](Full URL).
-- Have I avoided any placeholder text in my citations.
-- Can each link be clicked to access the original source.
+INSTRUCTIONS_END = """
+You have just executed tools and obtained results. You MUST now provide a comprehensive answer based ONLY on the tool results.
 \n\n\n
 """
 

src/engine/browser_engine.py

@@ -4,6 +4,7 @@
 #
 
 import requests
+from config import CONTENT_EXTRACTION, SEARCH_SELECTION
 from src.core.web_loader import web_loader
 
 class BrowserEngine:
@@ -58,7 +59,7 @@ class BrowserEngine:
             )
             request_response.raise_for_status()
             extracted_content = request_response.text
-            return f"{extracted_content}\n\n\n"
+            return f"{extracted_content}\n\n\n{CONTENT_EXTRACTION}\n\n\n"
         except Exception as error:
             return f"Error reading URL: {str(error)}"
 
@@ -82,6 +83,6 @@ class BrowserEngine:
             )
             search_response.raise_for_status()
             search_results = search_response.text
-            return f"{search_results}\n\n\n"
+            return f"{search_results}\n\n\n{SEARCH_SELECTION}\n\n\n"
         except Exception as error:
             return f"Error during search: {str(error)}"

src/processor/message_processor.py

@@ -4,7 +4,7 @@
 #
 
 import traceback
-from config import MODEL, INSTRUCTIONS
+from config import MODEL, INSTRUCTIONS_START
 from src.core.web_configuration import WebConfiguration
 from src.engine.browser_engine import BrowserEngine
 from src.tools.tool_manager import construct_tool_definitions
@@ -32,7 +32,7 @@ def process_user_request(user_message, chat_history):
         available_tools = construct_tool_definitions()
 
         conversation_messages = setup_response(
-            INSTRUCTIONS, 
+            INSTRUCTIONS_START, 
             chat_history, 
             user_message
         )

src/processor/response/generator.py

@@ -4,6 +4,7 @@
 #
 
 import traceback
+from config import INSTRUCTIONS_END
 
 def generate_response(
     server,
@@ -13,7 +14,14 @@ def generate_response(
     tools_done=False
 ):
     response_generator = ""
-
+    
+    if tools_done:
+        system_reminder = {
+            "role": "system",
+            "content": INSTRUCTIONS_END
+        }
+        conversation_messages.append(system_reminder)
+    
     try:
         response = server.chat.completions.create(
             model=model_name,

src/processor/response/setup.py

@@ -7,8 +7,8 @@ def setup_response(system_instruction, conversation_history, user_input):
     history = []
     
     if system_instruction:
-        history.insert(0, {"role": "system", "content": system_instruction})
-
+        history.append({"role": "system", "content": system_instruction})
+    
     if isinstance(conversation_history, list):
         for history_item in conversation_history:
             message_role = history_item.get("role")

src/processor/tools/interaction.py

@@ -13,7 +13,7 @@ from .executor import invoke_tool_function
 from config import REASONING_DELAY
 
 def process_tool_interactions(server, model_name, conversation_messages, tool_definitions, search_engine):
-    maximum_iterations = 3
+    maximum_iterations = 1
     logs_generator = ""
     tool_results = []