app/llm/async_llm_adapter.py

"""Async LLM adapter implementations used by the latency optimisation pipeline."""

from __future__ import annotations

import logging
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING

from app.config import settings
from app.generator.prompts import (
    ENHANCE_SYSTEM_PROMPT,
    PROOFREAD_SYSTEM_PROMPT,
    RICS_PROMPT,
    VALIDATE_SYSTEM_PROMPT,
    build_enhance_prompt,
    build_lcel_invoke_vars,
    build_proofread_prompt,
    build_validate_prompt,
    max_context_tokens_for_survey_level,
    max_output_tokens_for_survey_level,
    top_p_for_ai_involvement,
)
from app.llm.llm_throttle import throttled_llm_call

logger = logging.getLogger(__name__)

if TYPE_CHECKING:
    from app.models.schemas import WritingStyleProfile


class AsyncLLMAdapter(ABC):
    @abstractmethod
    async def generate_section(
        self,
        skeleton: str,
        bullets: list[str],
        snippets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.2,
        creativity_hint: str = "",
        document_context: list[str] | None = None,
        style_anchor: str | None = None,
        hierarchy_section_snippets: list[str] | None = None,
        paragraph_snippets: list[str] | None = None,
        identity_facts: str | None = None,
        survey_level: int | None = None,
        reference_only_context: bool = False,
        ai_percent: int | None = None,
        interference_level: str | None = None,
    ) -> str: ...

    @abstractmethod
    async def proofread(
        self,
        text: str,
        bullets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.15,
        creativity_hint: str = "",
    ) -> str: ...

    @abstractmethod
    async def enhance(
        self,
        text: str,
        bullets: list[str],
        snippets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.2,
        creativity_hint: str = "",
    ) -> str: ...

    @abstractmethod
    async def validate_section_compliance(
        self,
        *,
        survey_level: int | None,
        section_code: str,
        bullets: list[str],
        evidence_snippets: list[str],
        text: str,
    ) -> str: ...

    @abstractmethod
    async def constrained_weave(
        self,
        *,
        section_code: str,
        section_title: str | None,
        bullets: list[str],
        standard_passages: list[str],
        survey_level: int | None = None,
        tenant_id: str | None = None,
    ) -> str: ...


class AsyncOpenAIAdapter(AsyncLLMAdapter):
    """Async generate mode via LangChain LCEL; proofread/enhance/validate via OpenAI ChatCompletions."""

    def __init__(self) -> None:
        from langchain_openai import ChatOpenAI
        from openai import AsyncOpenAI

        self._client = AsyncOpenAI(api_key=settings.openai_api_key)
        self._model = settings.chat_model
        self._lc_llm = ChatOpenAI(
            model=self._model,
            temperature=0.2,
            max_tokens=settings.max_output_tokens,
            api_key=settings.openai_api_key,
            max_retries=3,
        )

    async def _call_async(
        self,
        *,
        system: str,
        user: str,
        phase: str,
        section_id: str | None,
        max_tokens: int | None = None,
        temperature: float = 0.2,
        survey_level: int | None = None,
        interference_level: str | None = None,
        tenant_id: str | None = None,
    ) -> str:
        from app.llm.llm_throttle import make_cache_hit_slot
        from app.llm.prompt_cache import (
            build_chat_messages,
            log_openai_cache_usage,
            openai_extra_kwargs,
            prompt_caching_active,
        )

        cache_slot = make_cache_hit_slot()

        async def _do_call() -> str:
            messages = build_chat_messages(system=system, user=user)
            extra = openai_extra_kwargs(
                phase=phase,
                model=self._model,
                survey_level=survey_level,
                interference_level=interference_level,
                tenant_id=tenant_id,
            )
            response = await self._client.chat.completions.create(
                model=self._model,
                messages=messages,
                max_tokens=max_tokens or settings.max_output_tokens,
                temperature=temperature,
                **extra,
            )
            if prompt_caching_active():
                cache_slot[0] = log_openai_cache_usage(
                    response, phase=phase, section_id=section_id
                )
            return (response.choices[0].message.content or "").strip()

        return await throttled_llm_call(
            phase=phase,
            section_id=section_id,
            cache_hit_out=cache_slot,
            call=_do_call,
        )

    async def generate_section(
        self,
        skeleton: str,
        bullets: list[str],
        snippets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.2,
        creativity_hint: str = "",
        document_context: list[str] | None = None,
        style_anchor: str | None = None,
        hierarchy_section_snippets: list[str] | None = None,
        paragraph_snippets: list[str] | None = None,
        identity_facts: str | None = None,
        survey_level: int | None = None,
        reference_only_context: bool = False,
        ai_percent: int | None = None,
        interference_level: str | None = None,
        tenant_id: str | None = None,
    ) -> str:
        from langchain_core.output_parsers import StrOutputParser

        fine = paragraph_snippets if paragraph_snippets is not None else snippets

        out_tokens = max_output_tokens_for_survey_level(
            survey_level, interference_level=interference_level
        )
        ctx_tokens = max_context_tokens_for_survey_level(
            survey_level, interference_level=interference_level
        )
        vars_ = build_lcel_invoke_vars(
            skeleton=skeleton,
            bullets=bullets,
            snippets=None,
            max_context_tokens=ctx_tokens,
            style_profile=style_profile,
            creativity_hint=creativity_hint,
            document_snippets=document_context,
            section_snippets=None,
            hierarchy_section_snippets=hierarchy_section_snippets,
            paragraph_snippets=fine,
            style_anchor=style_anchor,
            identity_facts=identity_facts,
            survey_level=survey_level,
            reference_only_context=reference_only_context,
            ai_percent=ai_percent,
            interference_level=interference_level,
        )
        top_p = top_p_for_ai_involvement(ai_percent)
        chain = (
            RICS_PROMPT
            | self._lc_llm.bind(
                temperature=temperature, max_tokens=out_tokens, top_p=top_p
            )
            | StrOutputParser()
        )

        from app.llm.prompt_cache import (
            build_chat_messages,
            log_openai_cache_usage,
            openai_extra_kwargs,
            prompt_caching_active,
        )

        phase = "generate_section"

        if prompt_caching_active():
            system = str(vars_.get("system_content") or "")
            user = str(vars_.get("user_content") or "")
            extra = openai_extra_kwargs(
                phase=phase,
                model=self._model,
                survey_level=survey_level,
                interference_level=interference_level,
                tenant_id=tenant_id,
            )

            from app.llm.llm_throttle import make_cache_hit_slot

            cache_slot = make_cache_hit_slot()

            async def _cached_generate() -> str:
                response = await self._client.chat.completions.create(
                    model=self._model,
                    messages=build_chat_messages(system=system, user=user),
                    max_tokens=out_tokens,
                    temperature=temperature,
                    top_p=top_p,
                    **extra,
                )
                cache_slot[0] = log_openai_cache_usage(
                    response, phase=phase, section_id=None
                )
                return (response.choices[0].message.content or "").strip()

            return await throttled_llm_call(
                phase=phase,
                section_id=None,
                cache_hit_out=cache_slot,
                call=_cached_generate,
            )

        from app.llm.lcel_invoke import ainvoke_lcel_chain

        return await ainvoke_lcel_chain(
            chain,
            vars_,
            phase=phase,
            section_id=None,
        )

    async def proofread(
        self,
        text: str,
        bullets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.15,
        creativity_hint: str = "",
    ) -> str:
        from app.generator.prompts import build_proofread_prompt
        from app.chunking.splitter import count_tokens

        user_prompt = build_proofread_prompt(
            text=text,
            bullets=bullets,
            style_profile=style_profile,
            creativity_hint=creativity_hint,
        )
        out_cap = max(700, min(2400, count_tokens(text or "") + 200))
        return await self._call_async(
            system=PROOFREAD_SYSTEM_PROMPT,
            user=user_prompt,
            phase="proofread",
            section_id=None,
            max_tokens=out_cap,
            temperature=temperature,
        )

    async def enhance(
        self,
        text: str,
        bullets: list[str],
        snippets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.2,
        creativity_hint: str = "",
    ) -> str:
        from app.chunking.splitter import count_tokens

        user_prompt = build_enhance_prompt(
            text=text,
            bullets=bullets,
            snippets=snippets,
            max_context_tokens=settings.max_context_tokens,
            style_profile=style_profile,
            creativity_hint=creativity_hint,
        )
        out_cap = max(900, min(2400, count_tokens(text or "") + 800))
        return await self._call_async(
            system=ENHANCE_SYSTEM_PROMPT,
            user=user_prompt,
            phase="enhance",
            section_id=None,
            max_tokens=out_cap,
            temperature=temperature,
        )

    async def validate_section_compliance(
        self,
        *,
        survey_level: int | None,
        section_code: str,
        bullets: list[str],
        evidence_snippets: list[str],
        text: str,
    ) -> str:
        user_prompt = build_validate_prompt(
            survey_level=survey_level,
            section_code=section_code,
            bullets=bullets,
            evidence_snippets=evidence_snippets,
            text=text,
        )
        result = await self._call_async(
            system=VALIDATE_SYSTEM_PROMPT,
            user=user_prompt,
            phase="validate_section",
            section_id=section_code,
            max_tokens=220,
            temperature=0.0,
        )
        return (result or "").strip()

    async def constrained_weave(
        self,
        *,
        section_code: str,
        section_title: str | None,
        bullets: list[str],
        standard_passages: list[str],
        survey_level: int | None = None,
        tenant_id: str | None = None,
    ) -> str:
        from app.generator.prompts import _ASSEMBLY_SYSTEM_CORE

        cleaned_passages = [
            str(p).strip() for p in (standard_passages or []) if str(p).strip()
        ]
        cleaned_bullets = [str(b).strip() for b in (bullets or []) if str(b).strip()]
        if not cleaned_passages or not cleaned_bullets:
            return ""

        title_part = f" — {section_title}" if section_title else ""
        user = (
            f"SECTION: {section_code}{title_part}\n\n"
            "STANDARD SOURCE PASSAGES (preserve wording; weave NOTES facts into the slots):\n"
            + "\n".join(f"- {p}" for p in cleaned_passages)
            + "\n\nINSPECTOR'S RAW NOTES (substitute these specifics into the standards):\n"
            + "\n".join(f"- {b}" for b in cleaned_bullets)
            + "\n\nProduce the structurally-routed output now. Standard wording stays, "
            "note facts replace generic slots, no new sentences, no new claims."
        )

        from app.llm.llm_throttle import make_cache_hit_slot
        from app.llm.prompt_cache import (
            build_chat_messages,
            log_openai_cache_usage,
            openai_extra_kwargs,
            prompt_caching_active,
        )

        cache_slot = make_cache_hit_slot()
        phase = "constrained_weave"

        async def _do_call() -> str:
            msgs = build_chat_messages(system=_ASSEMBLY_SYSTEM_CORE, user=user)
            extra = openai_extra_kwargs(
                phase=phase,
                model=self._model,
                survey_level=survey_level,
                tenant_id=tenant_id,
            )
            response = await self._client.chat.completions.create(
                model=self._model,
                messages=msgs,
                max_tokens=600,
                temperature=0.0,
                top_p=0.1,
                **extra,
            )
            if prompt_caching_active():
                cache_slot[0] = log_openai_cache_usage(
                    response, phase=phase, section_id=section_code
                )
            return (response.choices[0].message.content or "").strip()

        return await throttled_llm_call(
            phase=phase,
            section_id=section_code,
            cache_hit_out=cache_slot,
            call=_do_call,
        )


class MockAsyncLLMAdapter(AsyncLLMAdapter):
    """Deterministic mock adapter for async paths without OpenAI."""

    def __init__(self) -> None:
        pass

    async def generate_section(
        self,
        skeleton: str,
        bullets: list[str],
        snippets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.2,
        creativity_hint: str = "",
        document_context: list[str] | None = None,
        style_anchor: str | None = None,
        hierarchy_section_snippets: list[str] | None = None,
        paragraph_snippets: list[str] | None = None,
        identity_facts: str | None = None,
        survey_level: int | None = None,
        reference_only_context: bool = False,
        ai_percent: int | None = None,
        interference_level: str | None = None,
    ) -> str:
        style_note = f" (style: {style_profile.tone})" if style_profile else ""
        summary = "; ".join(bullets[:3]) if bullets else "No facts provided"
        return f"Based on the available information{style_note}: {summary}."

    async def proofread(
        self,
        text: str,
        bullets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.15,
        creativity_hint: str = "",
    ) -> str:
        return (
            f"{text}\n---NOTES---\n"
            "No OpenAI key configured — proofreading not available in mock mode."
        )

    async def enhance(
        self,
        text: str,
        bullets: list[str],
        snippets: list[str],
        style_profile: "WritingStyleProfile | None" = None,
        temperature: float = 0.2,
        creativity_hint: str = "",
    ) -> str:
        extra = (
            f" Additional context from {len(snippets)} retrieved source(s) noted."
            if snippets
            else ""
        )
        return (
            f"{text}{extra} "
            "[No OpenAI key configured — full technical enhancement requires OPENAI_API_KEY.]"
        )

    async def validate_section_compliance(
        self,
        *,
        survey_level: int | None,
        section_code: str,
        bullets: list[str],
        evidence_snippets: list[str],
        text: str,
    ) -> str:
        return "PASS"

    async def constrained_weave(
        self,
        *,
        section_code: str,
        section_title: str | None,
        bullets: list[str],
        standard_passages: list[str],
        survey_level: int | None = None,
        tenant_id: str | None = None,
    ) -> str:
        return ""


_async_llm_adapter_instance: AsyncLLMAdapter | None = None


def get_async_llm_adapter() -> AsyncLLMAdapter:
    """Return a singleton async adapter (real OpenAI when key configured)."""
    global _async_llm_adapter_instance
    if _async_llm_adapter_instance is not None:
        return _async_llm_adapter_instance

    if settings.openai_api_key:
        _async_llm_adapter_instance = AsyncOpenAIAdapter()
    else:
        _async_llm_adapter_instance = MockAsyncLLMAdapter()
    return _async_llm_adapter_instance