Function process_single_sequence

scripts/run_web_thinker_report.py:433–872

Process a single sequence through its entire reasoning chain with MAX_TOKENS limit

(
    seq: Dict,
    client: AsyncOpenAI,
    aux_client: AsyncOpenAI,
    semaphore: asyncio.Semaphore,
    args: argparse.Namespace,
    search_cache: Dict,
    url_cache: Dict,
    batch_output_records: List[Dict],
)

async def process_single_sequence(
    seq: Dict,
    client: AsyncOpenAI,
    aux_client: AsyncOpenAI,
    semaphore: asyncio.Semaphore,
    args: argparse.Namespace,
    search_cache: Dict,
    url_cache: Dict,
    batch_output_records: List[Dict],
) -> Dict:
    """Process a single sequence through its entire reasoning chain with MAX_TOKENS limit"""
    
    # Initialize limits
    MAX_TOKENS = 50000
    MAX_INTERACTIONS = 80  # Maximum number of total interactions，应对复读
    total_interactions = 0  # Track total interactions

    # Generate search plan first
    print(f"Generating search plan...")
    question = seq['item']['Question']
    _, search_plan = await generate_response(
        client=aux_client,
        model_name=args.aux_model_name,
        prompt=get_search_plan_instruction(question),
        semaphore=semaphore,
        max_tokens=args.max_tokens // 2,
        bad_words=[f"{END_SEARCH_QUERY}{tokenizer.eos_token}"],
    )

    print(f"---Search plan:---\n{search_plan}")
    
    # Generate the full instruction with the plan
    user_prompt = get_report_webthinker_instruction(question, search_plan)
    seq['prompt'] = user_prompt
    
    # Initialize token counter with prompt tokens
    total_tokens = len(seq['prompt'].split())
    
    # Initialize web explorer interactions list and article-related variables
    seq['web_explorer'] = []
    article = ""
    summarized_article = ""
    document_memory = []  # Store all retrieved web page content
    
    # Initialize BM25 for document retrieval
    tokenized_docs = []
    bm25 = None
    
    # First response uses chat completion
    formatted_prompt, response = await generate_response(
        client=client,
        model_name=args.model_name,
        prompt=seq['prompt'],
        semaphore=semaphore,
        temperature=args.temperature,
        top_p=args.top_p,
        max_tokens=args.max_tokens,
        repetition_penalty=args.repetition_penalty,