compass_backoffice/src/agent/nodes/speech_enrichment_node.py

import time
import logging
import json
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.components.clients.speech_analytics_client import SpeechAnalyticsClient
from src.components.clients.exceptions.speech_exceptions import SpeechClientError
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

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.

    Each returned item has: protocolo, data_reclamacao, similaridade_pct,
    reasoning, plus the 7 detail keys (same as reclamacao_atual).
    """
    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 string from related items: one line per item with
    'Protocolo X (Y%): <reasoning>'. Empty list -> negative spec message.
    """
    if not related_items:
        return _NO_SIMILAR_MESSAGE

    lines = []
    for item in related_items:
        reasoning = (item.get("reasoning") or "").strip()
        if not reasoning:
            reasoning = "—"
        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:
    """
    Calls the LLM to score similarity between the current complaint description
    and each historical item. Returns the LLM-parsed JSON list. Items below
    SPEECH_SIMILARITY_THRESHOLD will be dropped later by _build_related_history.
    """
    if not current_complaint_description or not history_list:
        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.")
        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(f"Calling LLM for history similarity filtering. Items: {len(clean_history)}")

    # LLM generation telemetry is recorded by agent_framework.llm.providers.