code/src/agents/llm_runnable.py

"""
LangChain Runnable adapter for BaseLLM.

Wraps the custom BaseLLM so LangGraph's astream_events(version="v2") can
capture per-token deltas from the generate_answer node.

Providers that implement real stream() (Ollama, OpenAI/DeepSeek) yield genuine
tokens. vLLM falls back to yielding the full response as a single chunk.
"""

from __future__ import annotations

import asyncio
import logging
import os
from typing import Any, AsyncIterator, Iterator, Optional

from langchain_core.runnables import Runnable
from langchain_core.runnables.config import RunnableConfig

logger = logging.getLogger(__name__)


class BaseLLMRunnable(Runnable):
    """Expose BaseLLM as a LangChain Runnable for astream_events token capture.

    LangGraph's astream_events(version="v2") wraps our stream() generator
    automatically and emits on_chain_stream events for each yielded token —
    no manual callback management needed.
    """

    def __init__(self, llm: Any, **defaults: Any) -> None:
        self._llm = llm
        self._defaults = defaults

    # ── Sync paths ──────────────────────────────────────────────────────────

    def invoke(
        self,
        input: str,
        config: Optional[RunnableConfig] = None,
        **kwargs: Any,
    ) -> str:
        params = {**self._defaults, **kwargs}
        return self._llm.generate(user_prompt=input, **params)

    def stream(
        self,
        input: str,
        config: Optional[RunnableConfig] = None,
        **kwargs: Any,
    ) -> Iterator[str]:
        params = {**self._defaults, **kwargs}
        yield from self._llm.stream(user_prompt=input, **params)

    # ── Async paths ─────────────────────────────────────────────────────────

    async def ainvoke(
        self,
        input: str,
        config: Optional[RunnableConfig] = None,
        **kwargs: Any,
    ) -> str:
        params = {**self._defaults, **kwargs}
        return await asyncio.to_thread(self._llm.generate, user_prompt=input, **params)

    async def astream(
        self,
        input: str,
        config: Optional[RunnableConfig] = None,
        **kwargs: Any,
    ) -> AsyncIterator[str]:
        def _collect():
            return list(self.stream(input, config, **kwargs))

        tokens = await asyncio.get_event_loop().run_in_executor(None, _collect)
        for token in tokens:
            yield token


def get_chat_model(llm_client: Any) -> Any:
    """
    Return a LangChain BaseChatModel supporting .bind_tools() for ReAct use.

    Feature-detects provider from llm_client.get_model_info() and instantiates
    the appropriate LangChain chat model with temperature=0. Returns None if the
    provider has no LangChain adapter or if the import fails — callers fall back
    to the legacy hardcoded retrieve node in that case.
    """
    if llm_client is None:
        return None

    try:
        cfg = llm_client.get_model_info()
    except Exception:
        cfg = {}

    provider = (cfg.get("provider") or "").lower()
    model    = cfg.get("model_name") or ""
    base_url = cfg.get("base_url") or os.environ.get("LLM_BASE_URL", "")
    api_key  = cfg.get("api_key") or os.environ.get("LLM_API_KEY") or os.environ.get("OPENAI_API_KEY", "")

    try:
        if provider == "anthropic":
            from langchain_anthropic import ChatAnthropic
            return ChatAnthropic(model=model, temperature=0, api_key=api_key or None)

        if provider == "openai":
            from langchain_openai import ChatOpenAI
            kwargs: dict = {"model": model, "temperature": 0}
            if api_key:
                kwargs["api_key"] = api_key
            if base_url:
                kwargs["base_url"] = base_url
            # DeepSeek v4 thinking mode requires reasoning_content forwarded in every
            # follow-up turn — create_react_agent doesn't do that, causing 400 errors.
            # Use extra_body to disable thinking at the HTTP request level.
            if "deepseek" in base_url.lower() or "deepseek" in model.lower():
                kwargs["extra_body"] = {"thinking": {"type": "disabled"}}
            return ChatOpenAI(**kwargs)

    except ImportError as e:
        logger.warning("LangChain chat-model adapter unavailable for provider=%r: %s", provider, e)
    except Exception as e:
        logger.warning("Failed to build chat model for provider=%r: %s", provider, e)

    return None