compass_backoffice/src/agent/nodes/speech_enrichment_node.py

import time
import logging
import json
import re
from src.agent.state.agent_state import AgentState, update_state_metadata
from src.agent.state.steps import GraphStep
from src.agent.state.step_helpers import set_current_step
from src.utils.observer import trace_node, score_current_trace
from src.utils.ics_collector import build_ic_payload, build_noc_api_metadata, build_noc_llm_metadata
from src.compat.framework_observer import event
from src.providers.llm_provider import classification_llm, chat_llm_with_usage, LLM_ENDPOINT
from src.core.config import settings
from src.core.prompt_manager import get_prompt
from src.agent.local_prompts.speech_history_analysis import speech_history_similarity_pt
from src.compat.framework_services import call_mcp_tool

logger = logging.getLogger(__name__)

_NO_SIMILAR_MESSAGE = "Não foram encontradas reclamações anteriores similares à reclamação atual"
_HISTORY_UNAVAILABLE_MESSAGE = "Indisponível SPEECH para Histórico de Reclamação - realizar consulta manualmente"

_HISTORY_DETAIL_KEYS = (
    "reclamacao_resumo",
    "causa_raiz",
    "descortesia_cliente",
    "motivo_reclamacao",
    "submotivo_reclamacao",
    "sentimento_cliente",
    "solucao_proposta_cliente",
)

_NULL_SPEECH_DATA = {
    "reclamacao_resumo": None,
    "causa_raiz": None,
    "descortesia_cliente": None,
    "motivo_reclamacao": None,
    "submotivo_reclamacao": None,
    "sentimento_cliente": None,
    "solucao_proposta_cliente": None,
    "analise_agente": None,
}


def _map_prediction_response(response: dict) -> dict:
    """Map the raw Prediction API response to the internal speech_analytics schema."""
    variables = response.get("variables", {})
    return {
        "reclamacao_resumo": response.get("resume") or variables.get("resume"),
        "causa_raiz": variables.get("causa_raiz"),
        "descortesia_cliente": variables.get("descortesia_cliente"),
        "motivo_reclamacao": variables.get("motivo"),
        "submotivo_reclamacao": variables.get("submotivo"),
        "sentimento_cliente": variables.get("sentimento_cliente"),
        "solucao_proposta_cliente": variables.get("solucao_proposta_cliente"),
    }


def _build_related_history(raw_history: list, current_complaint_id: str | None, llm_scores: list, threshold: int) -> list:
    """Merge LLM similarity scores with raw history items and filter by threshold."""
    if not raw_history or not llm_scores:
        return []

    history_by_protocol = {str(item.get("protocolo")): item for item in raw_history}
    related: list[dict] = []

    for score in llm_scores:
        protocolo = str(score.get("protocolo")) if score.get("protocolo") is not None else None
        if not protocolo or protocolo == str(current_complaint_id or ""):
            continue
        try:
            similaridade_pct = int(score.get("similaridade_pct", 0))
        except (TypeError, ValueError):
            continue
        if similaridade_pct < threshold:
            continue

        raw_item = history_by_protocol.get(protocolo)
        if not raw_item:
            continue

        merged = {
            "protocolo": raw_item.get("protocolo"),
            "data_reclamacao": raw_item.get("data_reclamacao"),
            "similaridade_pct": similaridade_pct,
            "reasoning": (score.get("reasoning") or "").strip(),
        }
        for key in _HISTORY_DETAIL_KEYS:
            merged[key] = raw_item.get(key)
        related.append(merged)

    related.sort(key=lambda x: x["similaridade_pct"], reverse=True)
    return related


def _build_analise_agente(related_items: list) -> str:
    """Build the analise_agente text from related items."""
    if not related_items:
        return _NO_SIMILAR_MESSAGE

    lines = []
    for item in related_items:
        reasoning = (item.get("reasoning") or "").strip() or "—"
        lines.append(f"Protocolo {item.get('protocolo')} ({item.get('similaridade_pct')}%): {reasoning}")
    return "\n".join(lines)


async def _score_history_with_llm(
    current_id: str | None,
    current_complaint_description: str,
    history_list: list,
    session_id: str = "",
    state: dict | None = None,
) -> list:
    """Score similarity between current complaint and historical complaints using the LLM."""
    if not current_complaint_description or not history_list:
        event("AGA.015", {
            "status": "Houve acionamento do Agente Especializado (LLM): não — sem resumo atual ou histórico vazio",
            "type": "INFO",
            "session_id": session_id,
            "tag": "AGA.015",
            "call_id": session_id,
            "origin": "LLM",
            **build_ic_payload({"session_id": session_id}, "AGA.015"),
        })
        return []

    clean_history = [item for item in history_list if str(item.get("protocolo")) != str(current_id or "")]
    if not clean_history:
        logger.info("No historical items left after deduplication.")
        event("AGA.015", {
            "status": "Houve acionamento do Agente Especializado (LLM): não — histórico vazio após deduplicação",
            "type": "INFO",
            "session_id": session_id,
            "tag": "AGA.015",
            "call_id": session_id,
            "origin": "LLM",
            **build_ic_payload({"session_id": session_id}, "AGA.015"),
        })
        return []

    llm = classification_llm
    prompt_template = get_prompt("speech_history_similarity_pt", speech_history_similarity_pt)

    history_payload = [
        {
            "protocolo": item.get("protocolo"),
            "motivo_reclamacao": item.get("motivo_reclamacao"),
            "submotivo_reclamacao": item.get("submotivo_reclamacao"),
            "causa_raiz": item.get("causa_raiz"),
            "reclamacao_resumo": item.get("reclamacao_resumo"),
        }
        for item in clean_history
    ]

    message = prompt_template.format(
        current_complaint_description=current_complaint_description,
        history_json=json.dumps(history_payload, ensure_ascii=False),
    )

    logger.info("Calling LLM for history similarity filtering. Items: %s", len(clean_history))
    event("AGA.014", {
        "status": "Houve acionamento do Agente Especializado (LLM): sim — filtragem de histórico Speech",
        "type": "INFO",
        "session_id": session_id,
        "tag": "AGA.014",
        "call_id": session_id,
        "origin": "LLM",
        **build_ic_payload({"session_id": session_id}, "AGA.014"),
    })

    try:
        _t0 = time.perf_counter()
        llm_resp = chat_llm_with_usage(llm, message)
        content = llm_resp.content

        json_match = re.search(r"```(?:json)?\s*(\[.*?\])\s*```", content, re.DOTALL)
        if json_match:
            content = json_match.group(1)
        else:
            json_match = re.search(r"(\[.*?\])", content, re.DOTALL)
            if json_match:
                content = json_match.group(1)

        scores = json.loads(content)
        if not isinstance(scores, list):
            raise ValueError(f"LLM returned non-list: {type(scores).__name__}")

        score_current_trace(name="speech_history_similarity_valid", value=1.0)
        return scores
    except Exception as e:
        logger.error("Error during LLM history scoring: %s", e, exc_info=True)
        score_current_trace(name="speech_history_similarity_valid", value=0.0, comment=str(e))
        if state is not None:
            event("NOC.004", {
                "status": f"LLM error during history similarity analysis: {e}",
                "type": "FAILURE",
                **build_noc_llm_metadata(
                    state,
                    "NOC.004",
                    latency_ms=int((time.perf_counter() - _t0) * 1000) if "_t0" in locals() else 0,
                    llm_endpoint=LLM_ENDPOINT,
                    model_name=str(getattr(llm, "eligibleModel_name", "unknown")),
                ),
            }, metadata={"noc": True})
        return []


@trace_node
async def enrich_with_speech(state: AgentState) -> AgentState:
    """Enrich the ticket with Speech Analytics NLP insights and historical similarity."""
    logger.info("Running speech enrichment node.")

    await set_current_step(state, GraphStep.SPEECH_ENRICHMENT)
    session_id = state.get("session_id", "")
    context = state.get("metadata", {}).get("request_context", {})

    if "speech_analytics" in context:
        logger.info("Speech analytics data already exists in context (cached). Skipping API call.")
        return state

    complaint = context.get("complaint", {})
    reclamacao_id: str | None = complaint.get("complaintProtocol")
    raw_text: str | None = complaint.get("description")

    customer = context.get("customer", {})
    cpf_cnpj: str | None = customer.get("cpfCnpj")
    msisdn: str | None = customer.get("msisdn")

    # Framework-native path: call Speech through MCP/fallback instead of direct legacy clients.
    mcp_response = await call_mcp_tool(
        "consultar_speech_analytics",
        {
            "protocol_id": reclamacao_id,
            "customer_key": cpf_cnpj or msisdn,
            "interaction_key": reclamacao_id,
            "document": cpf_cnpj,
        },
        business_context={"customer_key": cpf_cnpj or msisdn, "interaction_key": reclamacao_id},
        original_context=context,
    )
    mcp_result = mcp_response.get("result") if isinstance(mcp_response.get("result"), dict) else mcp_response
    if mcp_response.get("ok") or str(mcp_result.get("source", "")).startswith("mock_"):
        summary = mcp_result.get("summary") or (raw_text or "Reclamação recebida para análise.")
        speech_data = {
            "reclamacao_resumo": summary,
            "causa_raiz": mcp_result.get("root_cause") or "Cobrança/contato indevido",
            "descortesia_cliente": mcp_result.get("rude_customer") or "False",
            "motivo_reclamacao": mcp_result.get("reason") or complaint.get("motive"),
            "submotivo_reclamacao": mcp_result.get("subreason") or complaint.get("modality"),
            "sentimento_cliente": mcp_result.get("sentiment") or "Negativo",
            "solucao_proposta_cliente": mcp_result.get("proposed_solution") or "Analisar cobrança e cessar contatos indevidos",
            "historico_bruto": mcp_result.get("events") or [],
            "historico_relacionado": mcp_result.get("related_history") or [],
            "analise_agente": mcp_result.get("agent_analysis") or (mcp_result.get("message") or _NO_SIMILAR_MESSAGE),
            "source": mcp_result.get("source", "framework_mcp"),
            "mcp_tool": "consultar_speech_analytics",
        }
        context = dict(context)
        context["speech_analytics"] = speech_data
        state = update_state_metadata(state, request_context=context)
        event("AGA.034", {
            "status": "Resultado Speech: sucesso via MCP/framework",
            "type": "SUCCESS",
            "session_id": session_id,
            "tag": "AGA.034",
            "call_id": session_id,
            "origin": "MCP",
            **build_ic_payload(state, "AGA.034"),
        }, metadata={"noc": True})
        return state

    if not getattr(settings, "BACKOFFICE_ALLOW_LEGACY_CLIENTS", False):
        speech_data = dict(_NULL_SPEECH_DATA)
        speech_data.update({
            "reclamacao_resumo": raw_text or "Indisponível SPEECH para Reclamação Atual - realizar consulta manualmente",
            "historico_bruto": [],
            "historico_relacionado": [],
            "analise_agente": _HISTORY_UNAVAILABLE_MESSAGE,
            "source": "framework_local_fallback",
        })
        context = dict(context)
        context["speech_analytics"] = speech_data
        state = update_state_metadata(state, request_context=context)
        await set_current_step(state, GraphStep.SPEECH_ENRICHMENT_UNAVAILABLE)
        return state

    from src.components.clients.speech_analytics_client import SpeechAnalyticsClient
    from src.components.clients.exceptions.speech_exceptions import SpeechClientError

    speech_data: dict = dict(_NULL_SPEECH_DATA)
    history_unavailable = False

    if not reclamacao_id or not raw_text:
        logger.warning(
            "Speech enrichment skipped prediction: missing complaintProtocol or description. "
            "reclamacao_id=%s, raw_text_present=%s",
            bool(reclamacao_id), bool(raw_text)
        )
        event("AGA.010", {
            "status": "Resultado consulta Speech: ignorada — complaintProtocol ou description ausente",
            "type": "INFO",
            "session_id": session_id,
            "tag": "AGA.010",
            "call_id": session_id,
            "origin": "AGENT",
            **build_ic_payload(state, "AGA.010"),
        }, metadata={"noc": True})
    else:
        try:
            client = SpeechAnalyticsClient()
            response, http_meta = await client.get_prediction(
                reclamacao_id=reclamacao_id,
                raw_text=raw_text,
                customer_segment="",
            )
            speech_data = _map_prediction_response(response)
            if not speech_data.get("reclamacao_resumo"):
                speech_data["reclamacao_resumo"] = "Reclamação Atual não foi encontrada no SPEECH"
                event("NOC.002", {
                    "status": "Invalid API response",
                    "type": "WARNING",
                    **build_noc_api_metadata(
                        state,
                        "NOC.002",
                        retry_count=http_meta.get("retry_count", 0),
                        latency_ms=http_meta["latency_ms"],
                        api_url=http_meta["url"],
                        status_code=http_meta["status_code"],
                    ),
                }, metadata={"noc": True})

            logger.info("Speech Analytics prediction retrieved successfully. reclamacao_id=%s", reclamacao_id)
            event("AGA.010", {
                "status": "Resultado consulta Speech: sucesso",
                "type": "SUCCESS",
                "session_id": session_id,
                "tag": "AGA.010",
                "call_id": session_id,
                "origin": "AGENT",
                **build_ic_payload(state, "AGA.010", {
                    "apiUrl": http_meta["url"],
                    "apiStatusCode": http_meta["status_code"],
                    "apiResponsePayload": http_meta["response_text"],
                    "latencyMs": http_meta["latency_ms"],
                }),
            }, metadata={"noc": True})
        except SpeechClientError as client_exc:
            logger.warning("Speech Analytics unavailable — continuing without prediction. Error: %s", client_exc)
            speech_data["reclamacao_resumo"] = "Indisponível SPEECH para Reclamação Atual - realizar consulta manualmente"
            await set_current_step(state, GraphStep.SPEECH_ENRICHMENT_UNAVAILABLE)
            api_fields = {
                "apiUrl": getattr(client_exc, "url", None) or "N/A",
                "apiStatusCode": getattr(client_exc, "status_code", None) if getattr(client_exc, "status_code", None) is not None else "N/A",
                "apiResponsePayload": getattr(client_exc, "response_text", None) if getattr(client_exc, "response_text", None) is not None else "N/A",
                "latencyMs": getattr(client_exc, "latency_ms", None) if getattr(client_exc, "latency_ms", None) is not None else "N/A",
            }
            event("AGA.035", {
                "status": f"Erro na consulta Speech (Reclamação Atual): SpeechClientError: {str(client_exc)}",
                "type": "FAILURE",
                "session_id": session_id,
                "tag": "AGA.035",
                "call_id": session_id,
                "origin": "AGENT",
                **build_ic_payload(state, "AGA.035", api_fields),
            }, metadata={"noc": True})

    historico: list = []
    if not cpf_cnpj:
        logger.warning("Speech enrichment skipped history fetch: cpfCnpj absent.", extra={"session_id": session_id})
        event("AGA.034", {
            "status": "Resultado consulta histórico Speech: ignorada — cpfCnpj ausente",
            "type": "INFO",
            "session_id": session_id,
            "tag": "AGA.034",
            "call_id": session_id,
            "origin": "AGENT",
            **build_ic_payload(state, "AGA.034"),
        }, metadata={"noc": True})
    else:
        try:
            client = SpeechAnalyticsClient()
            historico, http_meta = await client.get_history(cpf_cnpj, msisdn)
            logger.info("Speech Analytics history retrieved successfully. %s items.", len(historico), extra={"session_id": session_id})
            event("AGA.034", {
                "status": f"Resultado consulta histórico Speech: sucesso — {len(historico)} item(ns)",
                "type": "SUCCESS",
                "session_id": session_id,
                "tag": "AGA.034",
                "call_id": session_id,
                "origin": "AGENT",
                **build_ic_payload(state, "AGA.034", {
                    "apiUrl": http_meta["url"],
                    "apiStatusCode": http_meta["status_code"],
                    "apiResponsePayload": http_meta["response_text"],
                    "latencyMs": http_meta["latency_ms"],
                }),
            }, metadata={"noc": True})
            if not historico:
                event("NOC.002", {
                    "status": "Invalid API response",
                    "type": "WARNING",
                    **build_noc_api_metadata(
                        state,
                        "NOC.002",
                        retry_count=http_meta.get("retry_count", 0),
                        latency_ms=http_meta["latency_ms"],
                        api_url=http_meta["url"],
                        status_code=http_meta["status_code"],
                    ),
                }, metadata={"noc": True})
        except Exception as exc:
            logger.warning("Failed to fetch speech history: %s", exc)
            await set_current_step(state, GraphStep.SPEECH_ENRICHMENT_UNAVAILABLE)
            api_fields = {}
            if isinstance(exc, SpeechClientError):
                api_fields = {
                    "apiUrl": getattr(exc, "url", None) or "N/A",
                    "apiStatusCode": getattr(exc, "status_code", None) if getattr(exc, "status_code", None) is not None else "N/A",
                    "apiResponsePayload": getattr(exc, "response_text", None) if getattr(exc, "response_text", None) is not None else "N/A",
                    "latencyMs": getattr(exc, "latency_ms", None) if getattr(exc, "latency_ms", None) is not None else "N/A",
                }
            event("AGA.035", {
                "status": f"Falha ao buscar histórico Speech: {str(exc)}",
                "type": "FAILURE",
                "session_id": session_id,
                "tag": "AGA.035",
                "call_id": session_id,
                "origin": "AGENT",
                **build_ic_payload(state, "AGA.035", api_fields),
            }, metadata={"noc": True})
            history_unavailable = True

    historico_relacionado: list = []
    if historico and raw_text:
        scores = await _score_history_with_llm(
            current_id=reclamacao_id,
            current_complaint_description=raw_text,
            history_list=historico,
            session_id=session_id,
            state=state,
        )
        historico_relacionado = _build_related_history(
            raw_history=historico,
            current_complaint_id=reclamacao_id,
            llm_scores=scores,
            threshold=settings.SPEECH_SIMILARITY_THRESHOLD,
        )

    def _strip_to_detail_keys(h_list: list) -> list:
        return [{k: item.get(k) for k in _HISTORY_DETAIL_KEYS} for item in h_list]

    speech_data["historico_bruto"] = _strip_to_detail_keys(historico)
    speech_data["historico_relacionado"] = historico_relacionado
    speech_data["analise_agente"] = _HISTORY_UNAVAILABLE_MESSAGE if history_unavailable else _build_analise_agente(historico_relacionado)

    context = dict(context)
    context["speech_analytics"] = speech_data
    state = update_state_metadata(state, request_context=context)

    logger.info("Speech enrichment node completed successfully. State metadata and current step updated: %s", state.get("current_step"))
    return state