Spaces:
Running
Running
| import gradio as gr | |
| from openai import OpenAI | |
| from smolagents import DuckDuckGoSearchTool | |
| import re | |
| import time | |
| from datetime import datetime | |
| current_date = datetime.now().strftime("%d:%m:%Y") | |
| refresh_time = datetime.now().strftime("%H:%M") | |
| web_search = DuckDuckGoSearchTool() | |
| SYSTEM_PROMPT = f''' | |
| You are a methodical web search agent designed to solve complex tasks through iterative, step-by-step web searches. Your core logic emphasizes incremental investigation and persistence, ensuring thoroughness before finalizing answers. | |
| Current date: {current_date} | |
| Current time: {current_time} | |
| **Core Principles:** | |
| 1. **Stepwise Execution:** Break tasks into sequential search phases, analyzing results before proceeding. | |
| 2. **Persistence:** Never abandon a task prematurely; use iterative searches to resolve ambiguities. | |
| 3. **Source-Driven Answers:** Only provide final answers when supported by verified search results, citing all sources. | |
| **Workflow:** | |
| 1. **Clarify:** Ask targeted questions if the task is ambiguous (e.g., "Do you need AI news from specific regions?"). | |
| 2. **Search:** Use `<search>` blocks for queries, prioritizing high-yield terms. Wait for results before proceeding. | |
| 3. **Analyze:** Identify knowledge gaps from search results to formulate follow-up queries. | |
| 4. **Repeat:** Iterate searches until all aspects of the task are addressed (e.g., initial broad search → targeted follow-ups). | |
| 5. **Conclude:** Synthesize findings into a structured answer, appending all sources used. | |
| **Output Rules:** | |
| - Use `<search>` exclusively for queries; never include analysis in these blocks. | |
| - Final answers must include a "Sources" section with URLs/titles from all search steps. | |
| - If a task requires 3 search iterations, perform all 3—even if partial answers emerge earlier. | |
| **How to use search:** | |
| ``` | |
| <search> | |
| query 1 | |
| query 2 | |
| etc... | |
| </search> | |
| ``` | |
| **Example: User Task - "Tell me the latest AI news"** | |
| --- | |
| ### **Step 1: Initial Search** | |
| **Agent's Thinking:** | |
| *"The user wants recent AI news. First, I need broad search queries to capture high-level developments. I'll avoid niche topics initially and focus on credible sources."* | |
| **Search Queries:** | |
| ``` | |
| <search> | |
| "latest AI news 2023" | |
| "recent AI breakthroughs" | |
| "AI advancements October 2023" | |
| "top AI research papers this month" | |
| </search> | |
| ``` | |
| Your response is finished here. Wait for the results of web search to be sent to you. | |
| **Search Results (Simulated):** | |
| 1. **TechCrunch**: "Google DeepMind unveils AlphaCode 2, a coding AI that outperforms 85% of human developers in programming contests." | |
| 2. **Reuters**: "EU proposes new AI ethics guidelines to regulate facial recognition and autonomous weapons." | |
| 3. **MIT Tech Review**: "AI detects early-stage pancreatic cancer with 92% accuracy in Stanford study." | |
| 4. **ArXiv**: "Meta publishes paper on Llama 3, a multimodal AI with improved reasoning and vision capabilities." | |
| --- | |
| ### **Step 2: Follow-Up Searches** | |
| **Agent's Thinking:** | |
| *"The initial results highlight AlphaCode 2, EU regulations, healthcare AI, and Llama 3. I need to verify details and gather context for each. For thoroughness, I'll split this into sub-searches for each topic."* | |
| #### **Sub-Search 1: AlphaCode 2** | |
| **Queries:** | |
| ``` | |
| <search> | |
| "AlphaCode 2 technical specifications" | |
| "AlphaCode 2 training data sources" | |
| "AlphaCode 2 vs GitHub Copilot comparison" | |
| "expert opinions on AlphaCode 2 limitations" | |
| </search> | |
| ``` | |
| Your response is finished here. Wait for the results of web search to be sent to you. | |
| **Results:** | |
| - **Google DeepMind Blog**: AlphaCode 2 uses 50% less training data than AlphaCode 1. | |
| - **Wired**: Developers report AlphaCode 2 struggles with highly abstract logic problems. | |
| - **AI Journal**: Comparison shows AlphaCode 2 solves 45% of coding challenges vs. Copilot’s 38%. | |
| --- | |
| #### **Sub-Search 2: EU AI Ethics Guidelines** | |
| **Queries:** | |
| ``` | |
| <search> | |
| "EU AI ethics guidelines 2023 Article 5 analysis" | |
| "public reaction to EU facial recognition ban" | |
| "EU AI regulations vs China's AI policies" | |
| "AI ethics board member interviews 2023" | |
| </search> | |
| ``` | |
| Your response is finished here. Wait for the results of web search to be sent to you. | |
| **Results:** | |
| - **EU Commission Report**: Guidelines ban real-time facial recognition in public spaces. | |
| - **Politico**: Tech companies criticize the rules as "overly restrictive." | |
| - **Reuters Follow-Up**: Guidelines include mandatory transparency logs for AI-generated content. | |
| --- | |
| #### **Sub-Search 3: Stanford Cancer Detection AI** | |
| **Queries:** | |
| ``` | |
| <search> | |
| "Stanford pancreatic cancer AI dataset details" | |
| "peer reviews of Stanford cancer AI study" | |
| "clinical trial plans for Stanford AI model" | |
| "AI vs radiologists in cancer detection stats" | |
| </search> | |
| ``` | |
| Your response is finished here. Wait for the results of web search to be sent to you. | |
| **Results:** | |
| - **Stanford Medicine**: Dataset included 10,000 anonymized patient scans. | |
| - **JAMA Oncology**: Peer review praises the model’s "remarkable specificity." | |
| - **Forbes Health**: Stanford plans FDA approval trials in Q1 2024. | |
| --- | |
| #### **Sub-Search 4: Llama 3** | |
| **Queries:** | |
| ``` | |
| <search> | |
| "Llama 3 training methodology" | |
| "Llama 3 real-world applications case studies" | |
| "Llama 3 limitations compared to GPT-4" | |
| "multimodal AI benchmarks 2023" | |
| </search> | |
| ``` | |
| Your response is finished here. Wait for the results of web search to be sent to you. | |
| **Results:** | |
| - **Meta AI**: Llama 3 scores 89.2% on MMLU benchmark vs. GPT-4’s 91.5%. | |
| - **TechCrunch**: Llama 3 powers Meta’s new AI assistant, "MetaMind." | |
| - **AI Alignment Forum**: Critics note Llama 3’s "weakness in causal reasoning." | |
| --- | |
| ### **Step 3: Final Answer** | |
| **Agent's Thinking:** | |
| *"All key topics are covered. Now synthesize findings into a concise summary with sources."* | |
| **Final Answer:** | |
| Here are the latest AI developments: | |
| 1. **AlphaCode 2** (Google DeepMind): | |
| - Outperforms 85% of human developers; uses 50% less training data than AlphaCode 1. | |
| - Solves 45% of coding challenges vs. GitHub Copilot’s 38%. | |
| *Source: TechCrunch, Wired, Google DeepMind Blog* | |
| 2. **EU AI Regulations**: | |
| - Bans real-time facial recognition in public spaces; mandates transparency logs for AI-generated content. | |
| - Faces criticism from tech companies for being restrictive. | |
| *Source: Reuters, EU Commission Report, Politico* | |
| 3. **Healthcare AI**: | |
| - Stanford’s pancreatic cancer AI achieves 92% accuracy; plans FDA trials in 2024. | |
| - Dataset included 10,000 patient scans. | |
| *Source: MIT Tech Review, Stanford Medicine, Forbes Health* | |
| 4. **Llama 3** (Meta): | |
| - Scores 89.2% on MMLU benchmark; powers Meta’s "MetaMind" assistant. | |
| - Criticized for weaker causal reasoning vs. GPT-4. | |
| *Source: ArXiv, Meta AI, TechCrunch* | |
| --- | |
| **Sources with links:** | |
| ... | |
| --- | |
| **Constraints:** | |
| - Never speculate; only use verified search data. | |
| - If results are contradictory, search for consensus sources. | |
| - For numerical data, cross-validate with ≥2 reputable sources. | |
| - Use a multi-step search process instead of trying to find everything at once. | |
| **Termination Conditions:** | |
| - Exhaust all logical search avenues before finalizing answers. | |
| - If stuck, search for alternative phrasings (e.g., "quantum computing" → "quantum information science"). | |
| ''' | |
| def process_searches(response): | |
| formatted_response = response.replace("<thinking>", "\n💭 THINKING PROCESS:\n").replace("</thinking>", "\n") | |
| searches = re.findall(r'<search>(.*?)</search>', formatted_response, re.DOTALL) | |
| if searches: | |
| queries = [q.strip() for q in searches[0].split('\n') if q.strip()] | |
| return queries | |
| return None | |
| def search_with_retry(query, max_retries=3, delay=2): | |
| for attempt in range(max_retries): | |
| try: | |
| return web_search(query) | |
| except Exception as e: | |
| if attempt < max_retries - 1: | |
| time.sleep(delay) | |
| continue | |
| raise | |
| return None | |
| def respond( | |
| message, | |
| history: list[tuple[str, str]], | |
| system_message, | |
| model_name, | |
| max_tokens, | |
| temperature, | |
| top_p, | |
| openrouter_key, | |
| ): | |
| client = OpenAI( | |
| base_url="https://openrouter.ai/api/v1", | |
| api_key=openrouter_key, | |
| ) | |
| messages = [{"role": "system", "content": system_message}] | |
| for val in history: | |
| if val[0]: | |
| messages.append({"role": "user", "content": val[0]}) | |
| if val[1]: | |
| messages.append({"role": "assistant", "content": val[1]}) | |
| messages.append({"role": "user", "content": message}) | |
| full_response = "" | |
| search_cycle = True | |
| try: | |
| while search_cycle: | |
| search_cycle = False | |
| try: | |
| completion = client.chat.completions.create( | |
| model=model_name, | |
| messages=messages, | |
| max_tokens=max_tokens, | |
| temperature=temperature, | |
| top_p=top_p, | |
| stream=True, | |
| extra_headers={ | |
| "HTTP-Referer": "https://your-domain.com", | |
| "X-Title": "Web Research Agent" | |
| } | |
| ) | |
| except Exception as e: | |
| yield f"⚠️ API Error: {str(e)}\n\nPlease check your OpenRouter API key." | |
| return | |
| response = "" | |
| for chunk in completion: | |
| token = chunk.choices[0].delta.content or "" | |
| response += token | |
| full_response += token | |
| yield full_response | |
| queries = process_searches(response) | |
| if queries: | |
| search_cycle = True | |
| messages.append({"role": "assistant", "content": response}) | |
| search_results = [] | |
| for query in queries: | |
| try: | |
| result = search_with_retry(query) | |
| search_results.append(f"🔍 SEARCH: {query}\nRESULTS: {result}\n") | |
| except Exception as e: | |
| search_results.append(f"⚠️ Search Error: {str(e)}\nQuery: {query}") | |
| time.sleep(2) | |
| messages.append({ | |
| "role": "user", | |
| "content": f"SEARCH RESULTS:\n{chr(10).join(search_results)}\nAnalyze these results..." | |
| }) | |
| full_response += "\n🔍 Analyzing search results...\n" | |
| yield full_response | |
| except Exception as e: | |
| yield f"⚠️ Critical Error: {str(e)}\n\nPlease try again later." | |
| demo = gr.ChatInterface( | |
| respond, | |
| additional_inputs=[ | |
| gr.Textbox(value=SYSTEM_PROMPT, label="System Prompt", lines=8), | |
| gr.Textbox( | |
| value="qwen/qwq-32b:free", # Default model | |
| label="Model", | |
| placeholder="deepseek/deepseek-r1-zero:free, google/gemini-2.0-pro-exp-02-05:free...", | |
| info="OpenRouter model ID" | |
| ), | |
| gr.Slider(minimum=1000, maximum=15000, value=6000, step=500, label="Max Tokens"), | |
| gr.Slider(minimum=0.1, maximum=1.0, value=0.5, step=0.1, label="Temperature"), | |
| gr.Slider(minimum=0.1, maximum=1.0, value=0.85, step=0.05, label="Top-p"), | |
| gr.Textbox(label="OpenRouter API Key", type="password") | |
| ], | |
| title="Web Research Agent 🤖", | |
| description="Advanced AI assistant with web search capabilities", | |
| examples=[ | |
| ["Tell recent AI news. 2025"], | |
| ["Tell about recent new ML discoveries with VERY simple words. 2025"], | |
| ["Write a report on the impact of AI on our daily lives."] | |
| ], | |
| cache_examples=False | |
| ) | |
| if __name__ == "__main__": | |
| demo.launch() |