Ajustes conforme relatorio de testes 2026-08-27

2026-08-29 09:53:32 -03:00
parent 0ecff719b7
commit 88e1f070d7
791 changed files with 27040 additions and 29038 deletions
--- a/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/.env.example
+++ b/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/.env.example
@@ -14,45 +14,38 @@ CORS_ORIGINS=http://localhost:5173,http://127.0.0.1:5173
 # LLM - OCI Generative AI como provider principal
 ###############################################################################
 # Opções: mock, oci_openai, oci_sdk, openai_compatible
-LLM_PROVIDER=oci_sdk
+LLM_PROVIDER=oci_openai
 LLM_TEMPERATURE=0.2
 LLM_MAX_TOKENS=2048
 LLM_TIMEOUT_SECONDS=120

 # OCI OpenAI-compatible endpoint
-OCI_GENAI_BASE_URL=https://inference.generativeai.us-chicago-1.oci.oraclecloud.com
+OCI_GENAI_BASE_URL=https://inference.generativeai.us-chicago-1.oci.oraclecloud.com/openai/v1
 OCI_GENAI_MODEL=openai.gpt-4.1
-OCI_GENAI_API_KEY=sk-ph3FgX6iP3fxAQCXb9IpPIDTadkeeYAWntUWhzcWysIM6zsS
+OCI_GENAI_API_KEY=sk-ph3FgX6ph3FgX6ph3FgX6ph3FgX6ph3FgX6ph3FgX6
 OCI_GENAI_PROJECT_OCID=

-#OCI_GENAI_BASE_URL=https://pegruagntaiatenddev.pe.inference.generativeai.sa-saopaulo-1.oci.oraclecloud.com
-#OCI_GENAI_MODEL=openai.gpt-4.1
-#OCI_GENAI_API_KEY=
-#OCI_GENAI_PROJECT_OCID=
-
-
-# OCI_AUTH_MODE=config_file|instance_principal|resource_principal
-OCI_AUTH_MODE=config_file
 # OCI SDK / signer / profiles
 OCI_CONFIG_FILE=~/.oci/config
-OCI_PROFILE=LATINOAMERICA-Chicago
-OCI_COMPARTMENT_ID=ocid1.compartment.oc1..aaaaaaaaexpiw4a7dio64mkfv2t273s2hgdl6mgfvvyv7tycalnjlvpvfl3q
+OCI_PROFILE=DEFAULT
+OCI_COMPARTMENT_ID=ocid1.compartment.oc1..aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
 OCI_REGION=us-chicago-1

 ###############################################################################
 # Persistência
 ###############################################################################
 # Opções: memory, autonomous, mongodb
-SESSION_REPOSITORY_PROVIDER=autonomous
-MEMORY_REPOSITORY_PROVIDER=autonomous
-CHECKPOINT_REPOSITORY_PROVIDER=autonomous
+SESSION_REPOSITORY_PROVIDER=sqlite
+MEMORY_REPOSITORY_PROVIDER=sqlite
+CHECKPOINT_REPOSITORY_PROVIDER=sqlite
+SQLITE_DB_PATH=./data/agent_framework.db

 # Autonomous Database
 ADB_USER=admin
-ADB_PASSWORD=Moniquinha19721972
-ADB_DSN=oradb23ai_high
-ADB_WALLET_LOCATION=/mnt/d/Dropbox/ORACLE/LatinoAmerica/Wallet_ORADB23ai
-ADB_WALLET_PASSWORD=Moniquinha1972
+ADB_PASSWORD=fjhsdf04954hf
+ADB_DSN=oradb23aidev_high
+ADB_WALLET_LOCATION=/ORACLE/DEFAULT/Wallet_ORADB23aiDev
+ADB_WALLET_PASSWORD=fjhsdf04954hf
 ADB_TABLE_PREFIX=AGENTFW

 # MongoDB - também pode representar Autonomous usando API compatível com Mongo, se habilitada no ambiente
@@ -66,10 +59,10 @@ ENABLE_REDIS_CACHE=false
 ###############################################################################
 # RAG / Vector / Graph
 ###############################################################################
-VECTOR_STORE_PROVIDER=autonomous
-GRAPH_STORE_PROVIDER=autonomous
+VECTOR_STORE_PROVIDER=sqlite
+GRAPH_STORE_PROVIDER=sqlite
 RAG_TOP_K=5
-EMBEDDING_PROVIDER=oci
+EMBEDDING_PROVIDER=mock
 OCI_EMBEDDING_MODEL=cohere.embed-multilingual-v3.0
 RAG_FILE_GLOBS=*.md,*.txt,*.yaml,*.yml,*.json

@@ -77,21 +70,14 @@ RAG_FILE_GLOBS=*.md,*.txt,*.yaml,*.yml,*.json
 # Observabilidade
 ###############################################################################
 ENABLE_LANGFUSE=true
- # Opcional: verbose, compact
-LANGFUSE_TRACE_MODE=compact
-# Nome customizado do trace pai, ex.: backoffice.checklist.workflow ou backoffice.emulador.workflow
-LANGFUSE_COMPACT_VISIBLE_EVENT_PREFIXES=AGA.,NOC., IC.
-LANGFUSE_COMPACT_SUPPRESSED_PREFIXES=llm.chat_completion
-LANGFUSE_IGNORE_HEALTHCHECKS=true
-LANGFUSE_IGNORED_PATHS=/health,/ready,/metrics
-LANGFUSE_PUBLIC_KEY=pk-lf-4a1e3921-5158-4fd3-a16d-7a77549fb312
-LANGFUSE_SECRET_KEY=sk-lf-efc6fd59-c5ec-4858-b6ec-4aa129734915
+LANGFUSE_TRACE_MODE=compact # Opcional: verbose, compact
+LANGFUSE_PUBLIC_KEY=pk-lf-2f9da109-5b0f-4c78-b61d-9598ed787eba
+LANGFUSE_SECRET_KEY=sk-lf-a4cb0cdd-f2ea-4468-9911-cebeb91ba944
 LANGFUSE_HOST=http://localhost:3005
 ENABLE_OTEL=false
 OTEL_EXPORTER_OTLP_ENDPOINT=
 OTEL_SERVICE_NAME=ai-agent-template
 ENABLE_LANGFUSE_OPENAI_AUTO_INSTRUMENTATION=true
-ENABLE_LANGFUSE_ANALYTICS_PUBLISHER=false

 ###############################################################################
 # Analytics / Observer corporativo
@@ -99,7 +85,7 @@ ENABLE_LANGFUSE_ANALYTICS_PUBLISHER=false
 # Quando true, AgentObserver publica eventos IC.*, NOC.* e GRL.* nos providers abaixo.
 ENABLE_ANALYTICS=false
 # Providers aceitos: oci_streaming,pubsub,noop
-ANALYTICS_PROVIDERS=oci_streaming
+ANALYTICS_PROVIDERS=pubsub
 # Compatibilidade FIRST/TIM: pode informar AGENT_PUBSUB_TOPIC diretamente.
 AGENT_PUBSUB_TOPIC=
 GCP_PUBSUB_TOPIC_PATH=
@@ -162,6 +148,7 @@ ROUTE_STICKINESS_HISTORY_TURNS=2
 ROUTE_STICKINESS_MAX_TOKENS=80
 HUMAN_HANDOFF_MESSAGE=Vou encaminhar seu atendimento para uma pessoa.
 END_SESSION_MESSAGE=Atendimento encerrado. Obrigado pelo contato.
+SESSION_ALREADY_ENDED_MESSAGE=Este atendimento já foi encerrado. Inicie uma nova sessão para continuar.

 ###############################################################################
 # MCP / Tools
@@ -169,13 +156,14 @@ END_SESSION_MESSAGE=Atendimento encerrado. Obrigado pelo contato.
 ENABLE_MCP_TOOLS=true
 MCP_SERVERS_CONFIG_PATH=./config/mcp_servers.yaml
 TOOLS_CONFIG_PATH=./config/tools.yaml
+TOOL_POLICIES_PATH=./config/tool_policies.yaml
 MCP_TOOL_TIMEOUT_SECONDS=30

 # router = EnterpriseRouter seleciona um agente; supervisor = pode acionar múltiplos agentes
 ROUTING_MODE=router

 # Usage/cost accounting
-USAGE_REPOSITORY_PROVIDER=autonomous
+USAGE_REPOSITORY_PROVIDER=sqlite
 IDENTITY_CONFIG_PATH=./config/identity.yaml
 MCP_PARAMETER_MAPPING_PATH=./config/mcp_parameter_mapping.yaml

@@ -192,6 +180,18 @@ MEMORY_SUMMARY_USE_LLM=true
 MEMORY_INJECT_RECENT_MESSAGES=true
 MEMORY_INJECT_SUMMARY=true

+###############################################################################
+# MCP Gateway
+###############################################################################
+# true = framework routes tool calls to the dedicated MCP Gateway.
+# false = framework calls MCP servers directly from mcp_servers.yaml.
+MCP_GATEWAY_ENABLED=true
+MCP_GATEWAY_URL=http://localhost:8300
+MCP_GATEWAY_TIMEOUT_SECONDS=60
+# MCP_GATEWAY_TOKEN=
+MCP_GATEWAY_AGENT_ID=telecom_contas
+MCP_GATEWAY_TENANT_ID=default
+
 ###############################################################################
 # LONG-TERM MEMORY
 ###############################################################################
--- a/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/app/workflows/pycache/agent_graph.cpython-313.pyc
+++ b/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/app/workflows/pycache/agent_graph.cpython-313.pyc
--- a/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/app/workflows/agent_graph.py
+++ b/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/app/workflows/agent_graph.py
@@ -160,7 +160,7 @@ class AgentWorkflow:
        builder.add_conditional_edges(
            "input_guardrails",
            self._after_input_guardrails,
-            {"blocked": "persist", "continue": "load_long_term_memory"},
+            {"blocked": "output_guardrails", "continue": "load_long_term_memory"},
        )
        builder.add_edge("load_long_term_memory", "routing_decision")
        builder.add_conditional_edges(
@@ -197,6 +197,31 @@ class AgentWorkflow:
    def _after_input_guardrails(self, state):
        return "blocked" if state.get("blocked") else "continue"

+    @staticmethod
+    def _input_guardrail_user_message(decisions, state, sanitized_text):
+        # Keep the technical guardrail reason in telemetry, but expose only a
+        # safe, actionable message to the end user. The message is intentionally
+        # routed through output_guardrails before persistence/delivery.
+        blocked = [d for d in decisions if not getattr(d, "allowed", True)]
+        first = blocked[0] if blocked else None
+        code = str(getattr(first, "code", "") or "").upper()
+        if code == "COER":
+            return (
+                "Não consegui entender sua última mensagem porque ela parece "
+                "incompleta ou ambígua. Pode reformular ou completar o que você quis dizer?"
+            )
+        if code == "INPUT_SIZE":
+            return "Sua mensagem ficou muito longa para eu processar de uma vez. Pode resumir ou dividir em partes?"
+        if code == "DLEX_IN":
+            return "Não posso usar essa informação da forma solicitada. Reformule o pedido sem incluir dados ou conteúdo restrito."
+        if code == "PINJ":
+            return "Não posso seguir instruções que tentem alterar as regras do atendimento. Posso continuar ajudando com a sua solicitação."
+        if code == "TOX":
+            return "Não consegui prosseguir com essa mensagem. Pode reformular o pedido para continuarmos o atendimento?"
+        if code == "CMP":
+            return "Não posso prosseguir com essa solicitação dessa forma. Posso ajudar com uma alternativa permitida."
+        return "Não consegui processar essa mensagem. Pode reformular para eu continuar o atendimento?"
+
    async def input_guardrails(self, state):
        if state.get("session_ended") is True:
            answer = str(getattr(
@@ -281,12 +306,33 @@ class AgentWorkflow:
                component="workflow.input_guardrails.final",
            )
            if any(not d.allowed for d in decisions):
+                # A blocking input guardrail stops the turn before routing/tools.
+                # Clear turn-local routing/tool state so stale data from a prior
+                # turn cannot appear as if it was executed after the block.
+                user_message = self._input_guardrail_user_message(decisions, state, sanitized)
                return {
                    "sanitized_input": sanitized,
-                    "answer": "Não consegui seguir com essa mensagem por regra de segurança.",
-                    "final_answer": "Não consegui seguir com essa mensagem por regra de segurança.",
+                    "answer": user_message,
+                    "final_answer": None,
                    "guardrail_decisions": [d.model_dump() for d in decisions],
                    "route": "blocked",
+                    "intent": "input_guardrail_blocked",
+                    "route_decision": {
+                        "route": "blocked",
+                        "agent": None,
+                        "intent": "input_guardrail_blocked",
+                        "confidence": 1.0,
+                        "reason": "Entrada interrompida por guardrail antes do roteamento.",
+                        "method": "guardrail",
+                        "next_state": state.get("next_state"),
+                        "handoff": False,
+                        "metadata": {},
+                        "domain": state.get("domain"),
+                        "mcp_tools": [],
+                    },
+                    "mcp_tools": [],
+                    "mcp_results": [],
+                    "judge_results": [],
                    "blocked": True,
                }
            return {
--- a/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/config/routing.yaml
+++ b/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/config/routing.yaml
@@ -7,6 +7,35 @@ router:
  confidence_threshold: 0.65
  allow_handoff: true

+  transaction_confirmation:
+    # Explicit yes/no stays deterministic. Only inconclusive replies use this LLM fallback.
+    semantic_fallback:
+      enabled: true
+      allowed_values: [SIM, NAO, CONTINUAR]
+      confirm_values: [SIM]
+      reject_values: [NAO]
+      continue_values: [CONTINUAR]
+      include_relevant_context: true
+      profile_name: router
+      prompt: |
+        Você classifica a resposta do cliente a uma confirmação transacional pendente.
+        Considere a pergunta pendente, somente o histórico recente relacionado ao mesmo tema e a fala atual.
+        Não execute a ação e não invente fatos.
+
+        Classes permitidas: {{ allowed_values }}
+        - SIM: confirmação/aceite inequívoco, inclusive equivalentes como "isso mesmo", "pode confirmar", "é isso" quando o contexto tornar o aceite claro.
+        - NAO: recusa/cancelamento inequívoco da ação pendente.
+        - CONTINUAR: qualquer resposta que não confirme nem rejeite inequivocamente, incluindo pergunta adicional, correção, novo dado, ambiguidade ou possível mudança de assunto.
+
+        Pergunta pendente:
+        {{ pending_prompt }}
+
+        Histórico relevante:
+        {{ relevant_conversation_context }}
+
+        Resposta atual do cliente:
+        {{ user_input }}
+
 state_policies:
  - state: WAITING_BILLING_CONFIRMATION
    agent: billing_agent
--- a/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/docs/TRANSACTION_SEMANTIC_CONFIRMATION.md
+++ b/agent_framework_oci/Tuning-Performance/Transaction_Evidence/agent_template_backend/docs/TRANSACTION_SEMANTIC_CONFIRMATION.md
@@ -0,0 +1,11 @@
+# Confirmação Transacional Semântica
+
+Este template suporta confirmação transacional em duas camadas: primeiro um parser determinístico para `sim`/`não` e equivalentes explícitos; somente quando ele não consegue decidir, o framework usa um classificador semântico configurado em `config/routing.yaml`.
+
+A configuração `router.transaction_confirmation.semantic_fallback` usa três classes: `SIM`, `NAO` e `CONTINUAR`. O prompt pode usar `{{ pending_prompt }}`, `{{ relevant_conversation_context }}`, `{{ user_input }}` e `{{ allowed_values }}`. O histórico injetado é apenas contexto de interpretação; não substitui validação de negócio ou evidência MCP.
+
+Exemplo: após `Você confirma o cancelamento do serviço Tamboro Mensal?`, a frase `isso mesmo, pode confirmar` pode ser classificada como `SIM`. Já `mas qual é o valor?` deve ser `CONTINUAR`, portanto não executa a ação por confirmação.
+
+Entradas explícitas já suportadas continuam no caminho determinístico e não geram custo adicional de LLM. Em observabilidade, o fallback usa `transaction.confirmation.semantic_classifier` e o `route_decision.metadata` informa `transaction_confirmation_source: semantic`.
+
+Consulte `docs/developer/pt/03_transaction_workflows_and_state.md` do framework para o contrato completo e exemplos.