Agentes de IA no conversacionales mediante MLflow

Abrir la versión 'notebook' de esta página

Los agentes no conversacionales procesan entradas estructuradas para generar salidas específicas sin mantener el estado de la conversación. Cada solicitud es independiente y independiente, lo que hace que estos agentes sean ideales para operaciones específicas de tareas, como la clasificación de documentos, la extracción de datos, el análisis por lotes y la respuesta a preguntas estructuradas.

A diferencia de los agentes conversacionales que administran diálogos multiturno, los agentes no conversacionales se centran en ejecutar tareas claramente definidas de manera eficaz. Esta arquitectura simplificada permite un mayor rendimiento para las solicitudes independientes.

Aprenderá a:

Creación de un agente no conversacional
Implementación completa del seguimiento y la observabilidad de MLflow
Despliegue del agente en Model Serving con recopilación automática de trazas
Configuración del monitoreo de producción con los evaluadores de MLflow versión 3

Requisitos

Dependencias:

MLflow 3.2.0 o superior
databricks-agents 1.2.0 o superior
databricks-sdk[openai] para la integración de LLM
Python 3.10 o superior

Acceso al área de trabajo:

Acceso a las API de Foundation Model (valor predeterminado: Claude 3.7 Sonnet, configurable)
Acceso a un catálogo y un esquema para el registro del modelo de IA

%pip install --upgrade mlflow[databricks]==3.6.0 pydantic databricks-sdk[openai] databricks-agents databricks-sdk
%restart_python

Escenario de ejemplo

El agente de este ejemplo procesa preguntas estructuradas sobre el contenido del documento financiero y proporciona respuestas sí/no con razonamiento. Los usuarios proporcionan el texto del documento y las preguntas directamente en la entrada, lo que elimina la necesidad de infraestructura de búsqueda vectorial en este ejemplo simplificado. Esto muestra cómo los agentes no conversacionales pueden controlar tareas bien definidas sin contexto de conversación.

Puede ampliar este ejemplo para casos de uso de producción mediante la integración de herramientas y funcionalidades adicionales. Entre los ejemplos se incluyen la búsqueda de vectores para la recuperación de documentos, herramientas mcP (protocolo de contexto de modelo) para integraciones externas u otros agentes de Databricks, como Genie para el acceso a datos estructurados.

Configurar el principal de servicio

Los agentes no conversacionales no admiten la transferencia automática de autenticación para escribir rastros de Model Serving. En su lugar, debe implementar una integración personalizada de rastreo de MLflow 3 y gestionar la autenticación manualmente mediante una entidad de servicio.

Cree una entidad de servicio con credenciales de OAuth.
Almacenar credenciales en un ámbito de secreto:

# TODO: Configuration constants - Update these for your environment
CATALOG = "main"
SCHEMA = "default"  # Replace with your schema name
SECRET_SCOPE = "<YOUR_SECRET_SCOPE>"  # Replace with your secret scope name
DATABRICKS_HOST = (
    "https://host.databricks.com"  # Replace with your workspace URL
)

# TODO: If you have not yet stored your service principal's OAuth client id and client secret as Databricks secrets,
# uncomment the following code and replace the <client_id> and <client_secret> with your service principal's id and secret.

# from databricks.sdk import WorkspaceClient

# w = WorkspaceClient()
# w.secrets.put_secret(SECRET_SCOPE, "client_id", string_value ="<YOUR_SERVICE_PRINCIPAL_CLIENT_ID>")
# w.secrets.put_secret(SECRET_SCOPE, "client_secret", string_value ="<YOUR_SERVICE_PRINCIPAL_CLIENT_SECRET>")

Configure el experimento de MLflow:

Cree el experimento si no existe.
Conceda a la entidad de servicio CAN_EDIT permisos para el experimento.

# Mlflow experiment to capture traces
EXPERIMENT_NAME = "/Workspace/Shared/non-conversational"

# LLM Configuration
LLM_MODEL = "databricks-claude-3-7-sonnet"  # Change this to use different models

# Model and endpoint names - do not need to be changed
MODEL_NAME = "document_analyser"
ENDPOINT_NAME = "document_analyser_agent"
REGISTERED_MODEL_NAME = f"{CATALOG}.{SCHEMA}.{MODEL_NAME}"

from databricks.sdk import WorkspaceClient
from databricks.sdk.service.ml import ExperimentAccessControlRequest
from databricks.sdk.service.iam import PermissionLevel
import mlflow

# Set experiment and get the experiment object directly
experiment = mlflow.set_experiment(EXPERIMENT_NAME)
experiment_id = experiment.experiment_id

# Fetch the service principal client_id from secret scope
client_id = dbutils.secrets.get(scope=SECRET_SCOPE, key="client_id")

# Set permissions for the SPN which will later write the traces from the serving endpoint
w = WorkspaceClient()
# Set CAN_EDIT permissions for the service principal
w.experiments.set_permissions(
    experiment_id=experiment_id,
    access_control_list=[
        ExperimentAccessControlRequest(
            service_principal_name=client_id,
            permission_level=PermissionLevel.CAN_EDIT
        )
    ]
)

print(f"✓ CAN_EDIT permissions granted to SPN {client_id[:8]}... for experiment: {experiment_id}")

Formato de entrada y salida

A diferencia de los agentes conversacionales que usan formatos de mensajes de chat flexibles, los agentes no conversacionales requieren modelos pydantic estructurados para entradas y salidas:

Cree esquemas de entrada con todos los campos necesarios para la ejecución de tareas.
Incluya metadatos de seguimiento (trace_id, span_id) en esquemas de salida para habilitar el registro de comentarios.
Diseñe salidas que proporcionen explicaciones detalladas de razonamiento o cadena de pensamiento cuando proceda.
Valide los esquemas durante el desarrollo para detectar errores antes de la implementación.

Formato de entrada (`AgentInput`)

{
  "document_text": "Document content to analyze...",
  "questions": [
    { "text": "Do the documents contain a balance sheet?" },
    { "text": "Do the documents contain an income statement?" },
    { "text": "Do the documents contain a cash flow statement?" }
  ]
}

Formato de salida (`AgentOutput`)

{
  "results": [
    {
      "question_text": "Do the documents contain a balance sheet?",
      "answer": "Yes",
      "chain_of_thought": "Detailed reasoning for the answer...",
      "span_id": "abc123def456"
    }
  ],
  "trace_id": "tr-xyz789abc123"
}

Entrada estructurada: los usuarios proporcionan texto del documento y preguntas en una sola solicitud.
Razonamiento detallado: Cada respuesta incluye una cadena de pensamiento paso a paso
Rastreabilidad: la respuesta incluye trace_id y span_id para la recopilación de comentarios

Construir el agente no conversacional

Cree el agente no conversacional con el seguimiento de MLflow. El agente utiliza @mlflow.trace decoradores para capturar automáticamente las llamadas LLM y todo el flujo de solicitud, proporcionando así una observabilidad plena.

Los usuarios proporcionan el texto del documento y las preguntas directamente en la entrada.

%%writefile model.py
import json
import logging
from typing import Optional
import uuid
import os
import sys

from databricks.sdk import WorkspaceClient
import mlflow
from mlflow.pyfunc import PythonModel
from mlflow.tracing import set_destination
from mlflow.tracing.destination import Databricks
from mlflow.entities import SpanType

from pydantic import BaseModel, Field


class Question(BaseModel):
    """Represents a question in the input."""

    text: str = Field(..., description="Yes/no question about document content")


class AgentInput(BaseModel):
    """Input model for the document analyser agent."""

    document_text: str = Field(..., description="The document text to analyze")
    questions: list[Question] = Field(..., description="List of yes/no questions")


class Answer(BaseModel):
    """Represents a structured response from the LLM."""

    answer: str = Field(..., description="Yes or No answer")
    chain_of_thought: str = Field(..., description="Step-by-step reasoning for the answer")


class AnalysisResult(BaseModel):
    """Represents an analysis result in the output."""

    question_text: str = Field(..., description="Original question text")
    answer: str = Field(..., description="Yes or No answer")
    chain_of_thought: str = Field(..., description="Step-by-step reasoning for the answer")
    span_id: str | None = Field(None, description="MLflow span ID for this specific answer (None during offline evaluation)")


class AgentOutput(BaseModel):
    """Output model for the document analyser agent."""

    results: list[AnalysisResult] = Field(..., description="List of analysis results")
    trace_id: str | None = Field(None, description="MLflow trace ID for user feedback collection (None during offline evaluation)")


class DocumentAnalyser(PythonModel):
    """Non-conversational agent for document analysis using MLflow model serving."""

    def __init__(self) -> None:
        """Initialize the document analyser.

        Sets up logging configuration, initializes model properties, and prepares
        the model for serving.
        """
        self._setup_logging()
        self.model_name = "document_analyser_v1"
        self.logger.debug(f"Initialized {self.model_name}")

    def _setup_logging(self) -> None:
        """Set up logging configuration for Model Serving.

        Configures a logger that uses stderr for better visibility in Model Serving
        environments. Log level can be controlled via MODEL_LOG_LEVEL environment
        variable (defaults to INFO).
        """
        self.logger = logging.getLogger("ModelLogger")
        # Set log level from environment variable or default to INFO
        log_level = os.getenv("MODEL_LOG_LEVEL", "INFO").upper()
        self.logger.setLevel(getattr(logging, log_level, logging.INFO))
        if not self.logger.handlers:
            handler = logging.StreamHandler()
            handler.setLevel(getattr(logging, log_level, logging.INFO))
            formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
            handler.setFormatter(formatter)
            self.logger.addHandler(handler)

    def load_context(self, context) -> None:
        """Load model context and initialize clients.

        This method is called once when the model is loaded in the serving environment.
        It sets up MLflow tracing destination, initializes the Databricks workspace
        client, and configures the OpenAI-compatible client for LLM inference.

        Args:
            context: MLflow model context containing artifacts and configuration
        """
        self.logger.debug("Loading model context")
        set_destination(Databricks(experiment_id=os.getenv("MONITORING_EXPERIMENT_ID")))

        self.logger.debug("Instantiate workspace client")
        self.w = WorkspaceClient()
        # You can load any artifacts here if needed
        # self.artifacts = context.artifacts

        self.logger.debug("Instantiate openai client")
        # Get an OpenAI-compatible client configured for Databricks serving endpoints
        self.openai_client = self.w.serving_endpoints.get_open_ai_client()

    @mlflow.trace(name="answer_question", span_type=SpanType.LLM)
    def answer_question(self, question_text: str, document_text: str) -> tuple[object, str | None]:
        """Answer a question using LLM with structured response format.

        Uses the OpenAI-compatible client to call a language model with a structured
        JSON response format. The LLM analyzes the provided document text and returns
        a yes/no answer with reasoning.

        Args:
            question_text (str): The yes/no question to answer about the document
            document_text (str): The document text to analyze

        Returns:
            tuple: (openai.ChatCompletion, str|None) - LLM response and span_id
        """
        # Create a chat completion request with structured response for questions

        question_prompt = f"""
        You are a document analysis expert. Answer the following yes/no question based on the provided document.

        Question: "{question_text}"

        Document:
        {document_text}

        Analyze the document and provide a structured response.
        """

        # Create a separate sub-span for the actual OpenAI API call
        llm_response = self._call_openai_completion(question_prompt)

        # Get the current span ID for this specific answer
        current_span = mlflow.get_current_active_span()
        span_id = current_span.span_id if current_span is not None else None

        return llm_response, span_id

    @mlflow.trace(name="openai_completion", span_type=SpanType.LLM)
    def _call_openai_completion(self, prompt: str):
        """Make the actual OpenAI API call with its own sub-span.

        Args:
            prompt (str): The formatted prompt to send to the LLM

        Returns:
            OpenAI ChatCompletion response
        """
        return self.openai_client.chat.completions.create(
            model=os.getenv("LLM_MODEL", "databricks-claude-3-7-sonnet"),  # Configurable LLM model
            messages=[
                {
                    "role": "user",
                    "content": prompt
                }
            ],
            response_format={
                "type": "json_schema",
                "json_schema": {
                    "name": "question_response",
                    "schema": Answer.model_json_schema()
                }
            }
        )

    @mlflow.trace(name="document_analysis")
    def predict(self, context, model_input: list[AgentInput]) -> list[AgentOutput]:
        """Process document analysis questions with yes/no answers.

        Args:
            context: MLflow model context
            model_input: List of structured inputs containing document text and questions

        Returns:
            List of AgentOutput with yes/no answers and reasoning
        """
        self.logger.debug(f"Processing {len(model_input)} classification request(s)")

        # Get the current trace ID for user feedback collection
        # Will be None during offline evaluation when no active span exists
        current_span = mlflow.get_current_active_span()
        trace_id = current_span.trace_id if current_span is not None else None

        results = []
        for input_data in model_input:
            self.logger.debug(f"Number of questions: {len(input_data.questions)}")
            self.logger.debug(f"Document length: {len(input_data.document_text)} characters")

            analysis_results = []

            for question in input_data.questions:
                self.logger.debug(f"Processing question: {question.text}")

                # Answer the question using LLM with structured response
                llm_response, answer_span_id = self.answer_question(question.text, input_data.document_text)

                # Parse structured JSON response
                try:
                    response_data = json.loads(llm_response.choices[0].message.content)
                    answer_obj = Answer(**response_data)
                except Exception as e:
                    self.logger.debug(f"Failed to parse structured response: {e}")
                    # Fallback to default response
                    answer_obj = Answer(
                        answer="No",
                        chain_of_thought="Unable to process the question due to parsing error."
                    )

                analysis_results.append(AnalysisResult(
                    question_text=question.text,
                    answer=answer_obj.answer,
                    chain_of_thought=answer_obj.chain_of_thought,
                    span_id=answer_span_id
                ))

            self.logger.debug(f"Generated {len(analysis_results)} analysis results")

            results.append(AgentOutput(
                results=analysis_results,
                trace_id=trace_id
            ))

        return results

mlflow.models.set_model(DocumentAnalyser())

Registro y registro del agente

Para que el agente se pueda implementar en un punto de conexión de servicio, debe registrarse en un experimento de MLflow y registrarse en el catálogo de Unity.

import os
import mlflow

import json
from mlflow.pyfunc import PythonModel
from pydantic import BaseModel, Field
from model import DocumentAnalyser, AgentInput, Question


# Create example input for signature inference
def create_example_input() -> AgentInput:
    """Create example input for the non-conversational agent."""
    return AgentInput(
        document_text="Total assets: $2,300,000. Total liabilities: $1,200,000. Shareholder's equity: $1,100,000. Net income for the period was $450,000. Revenues: $1,700,000. Expenses: $1,250,000. Net cash provided by operating activities: $80,000. Cash flows from investing activities: -$20,000",
        questions=[
            Question(text="Do the documents contain a balance sheet?"),
            Question(text="Do the documents contain an income statement?"),
            Question(text="Do the documents contain a cash flow statement?"),
        ],
    )


input_example = create_example_input()

with mlflow.start_run(run_name="deploy_non_conversational_agent"):
    active_run = mlflow.active_run()
    current_experiment_id = active_run.info.experiment_id
    # Set environment variables for the model using current notebook experiment
    os.environ["MONITORING_EXPERIMENT_ID"] = current_experiment_id
    print(
        f"✓ Using current notebook experiment ID for tracing: {current_experiment_id}"
    )

    # Log the non-conversational agent with auto-inferred dependencies
    model_info = mlflow.pyfunc.log_model(
        name=MODEL_NAME,
        python_model="model.py",  # Path to the model code file
        input_example=[create_example_input().model_dump()],
        registered_model_name=REGISTERED_MODEL_NAME,
    )

# Set logged model as current active model to associate it with the below evaluation results
mlflow.set_active_model(model_id=mlflow.last_logged_model().model_id)

print(f"✓ Model logged and registered: {REGISTERED_MODEL_NAME}")
print(f"✓ Model version: {model_info.registered_model_version}")

Evaluación del agente

Antes de realizar la implementación en producción, evalúe el rendimiento del agente mediante el marco de evaluación de GenAI de MLflow con los puntuadores creados previamente. Algunos puntuadores requieren un conjunto de datos de referencia.

import mlflow
import mlflow.genai.datasets
from requests import HTTPError

# Create an evaluation dataset in Unity Catalog
uc_schema = f"{CATALOG}.{SCHEMA}"
evaluation_dataset_table_name = "document_analyser_eval"

try:
    # Try to create a new evaluation dataset
    eval_dataset = mlflow.genai.datasets.create_dataset(
        uc_table_name=f"{uc_schema}.{evaluation_dataset_table_name}",
    )
    print(f"✓ Created evaluation dataset: {uc_schema}.{evaluation_dataset_table_name}")
except HTTPError as e:
    # Check if it's a TABLE_ALREADY_EXISTS error
    if e.response.status_code == 400 and "TABLE_ALREADY_EXISTS" in str(e):
        print(
            f"Dataset {uc_schema}.{evaluation_dataset_table_name} already exists, loading existing dataset..."
        )
        eval_dataset = mlflow.genai.datasets.get_dataset(
            uc_table_name=f"{uc_schema}.{evaluation_dataset_table_name}"
        )
        print(
            f"✓ Loaded existing evaluation dataset: {uc_schema}.{evaluation_dataset_table_name}"
        )
    else:
        # Different HTTP error, re-raise
        raise

# Define comprehensive test cases with expected facts for ground truth comparison
sample_document = "Total assets: $2,300,000. Total liabilities: $1,200,000. Shareholder's equity: $1,100,000. Net income for the period was $450,000. Revenues: $1,700,000. Expenses: $1,250,000. Net cash provided by operating activities: $80,000. Cash flows from investing activities: -$20,000"

evaluation_examples = [
    {
        "inputs": {
            "document_text": sample_document,
            "questions": [{"text": "Do the documents contain a balance sheet?"}],
        },
        "expectations": {
            "expected_facts": [
                "answer is Yes",
                "balance sheet information",
                "total assets mentioned",
                "total liabilities mentioned",
                "shareholder's equity mentioned",
            ]
        },
    },
    {
        "inputs": {
            "document_text": sample_document,
            "questions": [{"text": "Do the documents contain an income statement?"}],
        },
        "expectations": {
            "expected_facts": [
                "answer is Yes",
                "income statement information",
                "net income mentioned",
                "revenues mentioned",
                "expenses mentioned",
            ]
        },
    },
    {
        "inputs": {
            "document_text": sample_document,
            "questions": [{"text": "Do the documents contain a cash flow statement?"}],
        },
        "expectations": {
            "expected_facts": [
                "answer is Yes",
                "cash flow information",
                "operating activities mentioned",
                "investing activities mentioned",
                "cash flows mentioned",
            ]
        },
    },
    {
        "inputs": {
            "document_text": sample_document,
            "questions": [
                {
                    "text": "Do the documents contain information about employee benefits?"
                }
            ],
        },
        "expectations": {
            "expected_facts": [
                "answer is No",
                "no employee benefits information",
                "financial statements focus",
                "no HR-related content",
            ]
        },
    },
]

# Add the examples to the evaluation dataset
eval_dataset.merge_records(evaluation_examples)
print(f"✓ Added {len(evaluation_examples)} records to evaluation dataset")

# Preview the dataset
df = eval_dataset.to_df()
print(f"✓ Dataset preview - Total records: {len(df)}")
df.display()

import warnings
import mlflow
from mlflow.genai.scorers import (
    RelevanceToQuery,
    Correctness,
    Guidelines,
)

# Suppress harmless threadpoolctl warnings that can appear in Databricks environments
warnings.filterwarnings("ignore", message=".*threadpoolctl.*")
warnings.filterwarnings("ignore", category=UserWarning, module="threadpoolctl")

# Load the logged model for evaluation
model_uri = f"models:/{REGISTERED_MODEL_NAME}/{model_info.registered_model_version}"
print(f"Loading model for evaluation: {model_uri}")

# Load the model as a predict function
loaded_model = mlflow.pyfunc.load_model(model_uri)


def my_app(document_text, questions):
    """Wrapper function for the model prediction."""
    # The evaluation dataset's inputs field contains {"document_text": "...", "questions": [...]}
    # but the predict_fn parameter names must match the keys in inputs
    input_data = {"document_text": document_text, "questions": questions}
    return loaded_model.predict([input_data])


# Define scorers for evaluation including ground truth comparison
correctness_scorer = Correctness()  # Compares against expected_facts
relevance_scorer = RelevanceToQuery()  # Evaluates relevance of response to question
response_schema_scorer = Guidelines(
    name="response_schema",
    guidelines="The response must be structured JSON with an 'answer' field containing 'Yes' or 'No' and a 'chain_of_thought' field with clear reasoning. There also needs to be a 'question_text' field that contains the question that was asked. All these fields are part of the 'results' array field.",
)  # Validates structured output format

# This creates an evaluation run using the MLflow-managed dataset
results = mlflow.genai.evaluate(
    data=eval_dataset,  # Use the MLflow-managed dataset
    predict_fn=my_app,
    scorers=[
        correctness_scorer,
        relevance_scorer,
        response_schema_scorer,
    ],
)

# Access the run ID
print(f"✓ Evaluation completed")
print(f"Evaluation run ID: {results.run_id}")

# Display evaluation results summary
if hasattr(results, "metrics") and results.metrics:
    print("\n📊 Evaluation Results Summary:")
    for metric_name, metric_value in results.metrics.items():
        if isinstance(metric_value, (int, float)):
            print(f"  • {metric_name}: {metric_value:.3f}")
        else:
            print(f"  • {metric_name}: {metric_value}")
else:
    print("✓ Evaluation completed - view detailed results in the evaluation experiment")

# Display link to the evaluation dataset
print(f"\n📊 Evaluation Dataset: {uc_schema}.{evaluation_dataset_table_name}")
print(f"🔗 View dataset in Unity Catalog Data Explorer")

Desplegar en Implementación del Modelo

Implemente el agente evaluado en un punto de conexión de servicio de modelos con las variables de entorno necesarias para el seguimiento de MLflow 3. Esto garantiza que todas las solicitudes de producción se realicen automáticamente un seguimiento y se registren en el experimento de MLflow especificado.

import mlflow
from databricks.sdk import WorkspaceClient
from databricks.sdk.service.serving import (
    ServedEntityInput,
    ServingModelWorkloadType,
    EndpointCoreConfigInput,
)

from model import DocumentAnalyser, AgentInput, Question

workspace = WorkspaceClient()

# Use the model version from the logged model
model_version = model_info.registered_model_version

print(f"Using model version: {model_version}")

new_entity = ServedEntityInput(
    entity_name=REGISTERED_MODEL_NAME,
    entity_version=model_version,
    name=f"{MODEL_NAME}-{model_version}",
    workload_size="Small",
    workload_type=ServingModelWorkloadType.CPU,
    scale_to_zero_enabled=True,
    environment_vars={
        "DATABRICKS_CLIENT_ID": f"{{{{secrets/{SECRET_SCOPE}/client_id}}}}",
        "DATABRICKS_CLIENT_SECRET": f"{{{{secrets/{SECRET_SCOPE}/client_secret}}}}",
        "DATABRICKS_HOST": DATABRICKS_HOST,
        "MLFLOW_TRACKING_URI": "databricks",
        "MONITORING_EXPERIMENT_ID": current_experiment_id,
        "MODEL_LOG_LEVEL": "INFO",
        "LLM_MODEL": LLM_MODEL,
    },
)

# Check if endpoint exists and create or update accordingly
try:
    # Try to get the existing endpoint
    existing_endpoint = workspace.serving_endpoints.get(ENDPOINT_NAME)
    print(
        f"Endpoint {ENDPOINT_NAME} exists, updating with model version {model_version}"
    )

    # Update existing endpoint with new model version
    workspace.serving_endpoints.update_config(
        name=ENDPOINT_NAME, served_entities=[new_entity]
    )
    print("Endpoint update initiated, waiting for completion...")

    # Wait for update to complete
    workspace.serving_endpoints.wait_get_serving_endpoint_not_updating(ENDPOINT_NAME)
    print("Endpoint updated successfully and is ready")

except Exception as e:
    # Endpoint doesn't exist, create it
    print(f"Endpoint {ENDPOINT_NAME} doesn't exist, creating new endpoint...")

    workspace.serving_endpoints.create(
        name=ENDPOINT_NAME,
        config=EndpointCoreConfigInput(name=ENDPOINT_NAME, served_entities=[new_entity]),
    )
    print("Endpoint creation initiated, waiting for completion...")

    # Wait for creation to complete
    workspace.serving_endpoints.wait_get_serving_endpoint_not_updating(ENDPOINT_NAME)
    print("Endpoint created successfully and is ready")

# Final status check
endpoint_status = workspace.serving_endpoints.get(ENDPOINT_NAME)
print(f"Final endpoint status: {endpoint_status.state}")
print(
    f"Endpoint URL: https://{DATABRICKS_HOST.replace('https://', '')}/serving-endpoints/{ENDPOINT_NAME}/invocations"
)

Configurar la monitorización de producción usando puntuadores

Configure la evaluación automática de la calidad para el tráfico de producción mediante los evaluadores de MLflow 3. Los puntuadores analizan automáticamente los seguimientos registrados desde las solicitudes de producción para proporcionar una supervisión continua de la calidad.

from mlflow.genai.scorers import (
    RelevanceToQuery,
    Guidelines,
    ScorerSamplingConfig,
    list_scorers,
    get_scorer,
)

# Set the active experiment for scoring (use the current notebook's experiment)
print(f"Setting experiment to: {current_experiment_id}")
mlflow.set_experiment(experiment_id=current_experiment_id)

# Verify the experiment is set correctly
current_experiment = mlflow.get_experiment(current_experiment_id)
print(
    f"Current experiment: {current_experiment.name} (ID: {current_experiment.experiment_id})"
)

# Setup scorers for production monitoring
print("Setting up production monitoring scorers...")

# Relevance scorer - always create new to avoid conflicts
relevance_scorer = RelevanceToQuery().register(name="financial_relevance_check")
relevance_scorer = relevance_scorer.start(
    sampling_config=ScorerSamplingConfig(sample_rate=0.5)
)
print("✅ Created relevance scorer (50% sampling)")

# Guidelines scorer for response schema validation
response_schema_scorer = Guidelines(
    name="response_schema",
    guidelines="The response must be structured JSON with an 'answer' field containing 'Yes' or 'No' and a 'chain_of_thought' field with clear reasoning.",
).register(name="response_schema_check")
response_schema_scorer = response_schema_scorer.start(
    sampling_config=ScorerSamplingConfig(sample_rate=0.4)
)
print("✅ Created response schema scorer (40% sampling)")

# List all active scorers
print(f"\nActive Scorers in Experiment {current_experiment_id}:")
scorers = list_scorers()
for scorer in scorers:
    print(f"• {scorer.name}: {scorer.sample_rate*100}% sampling")

Prueba el agente implementado

Pruebe el agente implementado con entradas de ejemplo. Cada solicitud generará automáticamente seguimientos de MLflow 3 que capturan el flujo de solicitud completo y los puntuadores de producción evaluarán estos seguimientos para la supervisión de calidad.

from databricks.sdk import WorkspaceClient

# Test the non-conversational agent endpoint using Databricks SDK
workspace = WorkspaceClient()

# Example payload with structured input for the non-conversational agent
test_input = {
    "inputs": [
        {
            "document_text": "Total assets: $2,300,000. Total liabilities: $1,200,000. Shareholder's equity: $1,100,000. Net income for the period was $450,000. Revenues: $1,700,000. Expenses: $1,250,000. Net cash provided by operating activities: $80,000. Cash flows from investing activities: -$20,000",
            "questions": [
                {"text": "Do the documents contain a balance sheet?"},
                {"text": "Do the documents contain an income statement?"},
                {"text": "Do the documents contain a cash flow statement?"},
            ],
        }
    ]
}

# Query the serving endpoint using the workspace client
response = workspace.serving_endpoints.query(
    name=ENDPOINT_NAME, inputs=test_input["inputs"]
)

print("Endpoint Response:")
print(response.as_dict())

# Generate MLflow experiment URL
experiment_url = f"{DATABRICKS_HOST}/ml/experiments/{current_experiment_id}"
print(f"\nMLflow Experiment URL: {experiment_url}")

Registrar comentarios de los usuarios

Incluso para los agentes no conversacionales, la recopilación de comentarios de los usuarios es fundamental para la mejora continua. Las aplicaciones front-end orientadas al usuario pueden permitir que los usuarios acepten o rechacen respuestas individuales proporcionadas por el agente. A continuación, se pueden registrar estos comentarios en MLflow utilizando el trace_id y el span_id incluidos en la respuesta.

Escenarios comunes de comentarios para agentes no conversacionales:

Comentarios de precisión: "¿Se ha corregido esta respuesta sí/no?"
Comentarios de relevancia: "¿Fue el razonamiento adecuado para la pregunta?"
Comentarios de calidad: "¿Fue suficiente la evidencia de apoyo?"
Informe de errores: "¿El agente ha entendido mal el contenido del documento?"

En la celda siguiente se muestra cómo registrar los comentarios de una respuesta individual mediante el span_id devuelto en la respuesta.

import mlflow
from mlflow.entities import AssessmentSource

# Get the response from the previous test (extract span_id from first result)
# In a real application, this would come from the API response
response_dict = response.as_dict()
first_prediction = response_dict["predictions"][0]
first_result = first_prediction["results"][0]

# Assert we have the required IDs for feedback logging
assert (
    first_result.get("span_id") is not None
), "span_id is required for feedback logging"
assert (
    first_prediction.get("trace_id") is not None
), "trace_id is required for feedback logging"

span_id = first_result["span_id"]
trace_id = first_prediction["trace_id"]
question_text = first_result["question_text"]
answer = first_result["answer"]

print(f"Logging feedback for question: '{question_text}'")
print(f"Agent answer: {answer}")
print(f"Span ID: {span_id}")
print(f"Trace ID: {trace_id}")

try:
    # Example: User provides positive feedback on this specific answer
    mlflow.log_feedback(
        trace_id=trace_id,
        span_id=span_id,
        name="user_feedback",
        value=True,  # True for positive, False for negative
        source=AssessmentSource(source_type="HUMAN"),
        rationale="Answer was accurate and well-reasoned",
    )

    print("✅ Feedback logged successfully!")

except Exception as e:
    print(f"Note: Could not log feedback in this environment: {e}")

Pasos siguientes

Aprenda a agregar herramientas al agente para ampliar las funcionalidades.
Revise la documentación de seguimiento de MLflow 3 para conocer las características de observabilidad avanzadas.
Documentación de supervisión de producción

Cuaderno de ejemplo

Agentes de IA no conversacionales mediante MLflow

Obtener el cuaderno

Comentarios

¿Le ha resultado útil esta página?

Last updated on 2025-11-20

Compartir a través de

Agentes de IA no conversacionales mediante MLflow

Requisitos

Escenario de ejemplo

Configurar el principal de servicio

Formato de entrada y salida

Formato de entrada (AgentInput)

Formato de salida (AgentOutput)

Construir el agente no conversacional

Registro y registro del agente

Evaluación del agente

Desplegar en Implementación del Modelo

Configurar la monitorización de producción usando puntuadores

Prueba el agente implementado

Registrar comentarios de los usuarios

Pasos siguientes

Cuaderno de ejemplo

Agentes de IA no conversacionales mediante MLflow

Comentarios

Recursos adicionales

Formato de entrada (`AgentInput`)

Formato de salida (`AgentOutput`)