Un problema persistente de los modelos de lenguaje grandes (LLM) es su tendencia a producir resultados impredecibles. A pesar de las indicaciones cuidadosamente elaboradas, los LLM a menudo se desvían de las expectativas, lo que hace que sus resultados sean difíciles de reutilizar o integrar en los flujos de trabajo. Peor aún, los resultados a menudo no son repetibles, lo que complica el procesamiento posterior.

En esta publicación del blog, exploramos la metaindicación estructurada , una técnica que genera dinámicamente esquemas JSON para soluciones antes de realizar tareas. Esto ayuda a crear indicaciones más reutilizables, confiables y predecibles.

Fondo

La mayoría de los LLM modernos ofrecen ahora un modo de salida JSON, pero los resultados no siempre se ajustan al esquema esperado. Los desarrolladores suelen recurrir a mecanismos de verificación y reintento, que consumen mucho tiempo, son costosos y tienden a afectar la experiencia del usuario.

Con gpt4o , OpenAI introdujo salidas estructuradas que garantizaban que la salida se ajustara a un esquema JSON proporcionado por el usuario. Aunque no se revelan los detalles exactos, es probable que esto se haya logrado mediante una combinación de decodificación restringida y sesgo logit previo a la muestra. Otras herramientas, como la guía, los esquemas, la inversión, CommandR y SGLang, ofrecen capacidades similares. Estos enfoques permiten a los desarrolladores guiar las salidas hacia esquemas JSON, cada uno con ventajas y desventajas únicas.

¿Qué es el meta-prompting estructurado?

La metaindicación estructurada es cualquier técnica que utiliza un LLM para crear dinámicamente una descripción estructurada de un problema antes de producir una solución. Este método ofrece varias ventajas con respecto a la inducción directa:

1. Dinámico : las estructuras de salida se generan en tiempo de ejecución según las descripciones de tareas, en lugar de estar codificadas o ajustadas en el modelo.
2. Adaptable : el esquema generado es legible para humanos y puede ser inspeccionado, verificado o modificado por humanos u otros LLM.
3. Reutilizable : el esquema se puede guardar y reutilizar en múltiples tareas, ejecuciones y máquinas.
4. Predecible : las soluciones se ajustarán a una estructura bien definida, lo que las hace adecuadas para el cálculo posterior.

Código

Veamos un ejemplo en el que utilizamos una metaindicación estructurada para crear un esquema para una nueva novela de suspenso y espías que se ha convertido en un éxito de ventas.

Todo el código para esta publicación de blog está disponible en este Colab Notebook .

1. Generar una estructura

El primer paso es definir un esquema JSON que describa el resultado deseado. Para ello, crearemos un esquema JSON para un esquema JSON (es decir, un metaesquema ). Esto se incluye como parte de la especificación, pero debemos realizar algunos cambios.

1.1 Ajuste del metaesquema

Las API de salida estructurada de OpenAI y Cohere imponen varias restricciones a los esquemas JSON estándar. Esto es bastante desafortunado, pero es algo que podemos solucionar por ahora. Modificaremos el metaesquema para garantizar la compatibilidad:

 from jsonschema import Draft202012Validator def openai_compatible_metaschema(schema: Dict[str, object]): schema["type"] = "object" del schema["allOf"] return schema openai_json_metaschema = openai_compatible_metaschema( copy.deepcopy(Draft202012Validator.META_SCHEMA) )

Nota *: Aparentemente* la inversión no tiene tales restricciones y admite esquemas JSON arbitrarios… ¡pero aún no tiene acceso público!

1.2 Agregar restricciones a nuestro meta-prompt

A continuación, incorporamos pautas en el mensaje para garantizar que el esquema JSON resultante cumpla con las restricciones de OpenAI. Por ejemplo:

• Todos los campos deben ser obligatorios.
• Los objetos tienen límites en cuanto a profundidad y tamaño de anidamiento.
• No se deben permitir propiedades adicionales ( "additionalProperties": false ).

 system_guidelines = "\n".join( [ "All fields must be required - To use Structured Outputs, all fields or function parameters must be specified as required. NOTE: Although all fields must be required (and the model will return a value for each parameter), it is possible to emulate an optional parameter by using a union type with null." "Objects have limitations on nesting depth and size - A schema may have up to 100 object properties total, with up to 5 levels of nesting.", "Limitations on total string size - In a schema, total string length of all property names, definition names, enum values, and const values cannot exceed 15,000 characters.", "Limitations on enum size - A schema may have up to 500 enum values across all enum properties. For a single enum property with string values, the total string length of all enum values cannot exceed 7,500 characters when there are more than 250 enum values.", "additionalProperties: false must always be set in objects - additionalProperties controls whether it is allowable for an object to contain additional keys / values that were not defined in the JSON Schema. Structured Outputs only supports generating specified keys / values, so we require developers to set additionalProperties: false to opt into Structured Outputs.", "Some type-specific keywords are not yet supported - Notable keywords not supported include: For strings: minLength, maxLength, pattern, format; For numbers: minimum, maximum, multipleOf; For objects: patternProperties, unevaluatedProperties, propertyNames, minProperties, maxProperties; For arrays: unevaluatedItems, contains, minContains, maxContains, minItems, maxItems, uniqueItems", "For anyOf, the nested schemas must each be a valid JSON Schema per this subset", "Definitions are supported - You can use definitions to define subschemas which are referenced throughout your schema. The following is a simple example.", "Recursive schemas are supported - Sample recursive schema using # to indicate root recursion.", ] )

1.3 Configuración del meta-prompt

Ahora definimos el meta-prompt para generar un esquema JSON para el esquema del thriller de espías.

 from langchain.prompts import ChatPromptTemplate task_description = "Write an outline for a bestselling spy thriller novel" task_guidelines = """ - You must follow one of the six basic story arcs: Rags to riches, Riches to rags, Icarus, Oedipus, Cinderella, Man in a hole - Outputs must include characters, plot points (including exposition, rising action, climax, falling action, and resolution), central conflict, setting, major turning points or "beats," character arcs, and a synopsis of the story; essentially, a detailed breakdown of the key elements that will drive the narrative throughout the novel. """ prompt_messages = ChatPromptTemplate.from_messages( [ ( "system", "You are an expert in creating JSON schemas. You have been asked to generate a detailed JSON schema for the output of a given task based based on a task desciption and some guidelines.", ), ( "system", "Your JSON schema must always adhere to the following system system guidelines for JSON schemas:\n<system_guidelines>\n{system_guidelines}\n</system_guidelines>", ), ( "user", "Use the task description and guidelines below to generate an output JSON schema for the following task based on the guidelines provided.\n\n<task_description>\n{task_description}\n</task_description>\n\n<guidelines>\n{task_guidelines}\n</guidelines>", ), ] )

1.4 Generando la estructura

Invocamos el LLM para crear el esquema:

 messages = prompt_messages.format_messages( system_guidelines=system_guidelines, task_description=task_description, task_guidelines=task_guidelines ) ## Make sure to set up OPENAI_API_KEY in your Colab Secrets ## https://x.com/GoogleColab/status/1719798406195867814 client = OpenAI(api_key=userdata.get('OPENAI_API_KEY')) model = "gpt-4o" metaprompt_completion = client.beta.chat.completions.parse( model=model, messages=convert_to_penai_messages(messages), response_format={ "type": "json_schema", "json_schema": JSONSchema( name="JsonMetaschema", description="JSON Metaschema for the 2020-12 Draft of the JSON Schema specification that can be used to validate JSON data", schema=openai_json_metaschema, strict=False, ) } ) task_output_schema = json.loads(metaprompt_completion.choices[0].message.content) print(json.dumps(task_output_schema, indent=2))

 { "$schema": "https://json-schema.org/draft/2020-12/schema", "title": "Outline for a Bestselling Spy Thriller Novel", "type": "object", "properties": { "storyArc": { "type": "string", "enum": [ "Rags to riches", "Riches to rags", "Icarus", "Oedipus", "Cinderella", "Man in a hole" ] }, "characters": { "type": "array", "items": { "type": "object", "properties": { "name": { "type": "string" }, "role": { "type": "string" }, "description": { "type": "string" }, "arc": { "type": "string" } }, "required": ["name", "role", "description", "arc"], "additionalProperties": false }, "minItems": 1 }, "plotPoints": { "type": "object", "properties": { "exposition": { "type": "string" }, "risingAction": { "type": "string" }, "climax": { "type": "string" }, "fallingAction": { "type": "string" }, "resolution": { "type": "string" } }, "required": [ "exposition", "risingAction", "climax", "fallingAction", "resolution" ], "additionalProperties": false }, "centralConflict": { "type": "string" }, "setting": { "type": "string" }, "majorTurningPoints": { "type": "array", "items": { "type": "string" }, "minItems": 1 }, "characterArcs": { "type": "object", "properties": { "protagonistArc": { "type": "string" }, "antagonistArc": { "type": "string" }, "supportingCharactersArcs": { "type": "array", "items": { "type": "string" }, "minItems": 0 } }, "required": [ "protagonistArc", "antagonistArc", "supportingCharactersArcs" ], "additionalProperties": false }, "synopsis": { "type": "string" } }, "required": [ "storyArc", "characters", "plotPoints", "centralConflict", "setting", "majorTurningPoints", "characterArcs", "synopsis" ], "additionalProperties": false }

Ahora tenemos un esquema que describe el esquema de nuestra novela de suspenso y espionaje. Este esquema se puede conservar en un archivo o en una base de datos.

2. Generar una solución

2.1 Configuración rápida

Utilizando el esquema, definimos un tema para el bosquejo de la novela. Usaremos algunas pautas e indicaciones básicas para el juego de roles:

 user_requirements = "Tell a story about counter-intelligence operative working against the clock. The novel should be extremely realistic, slow burn." task_prompt = ChatPromptTemplate.from_messages( [ ( "system", "You are a world-renowned author that has written dozens of bestselling thriller novels. Your task is to create an outline for a new novel based on the user's requirements.", ), ( "user", "Please write a novel outline based strictly on the following requirements <requirements>{requirements}</requirements>", ), ] ) task_completion = client.beta.chat.completions.parse( model=model, messages=convert_to_openai_messages(task_prompt.format_messages(requirements=user_requirements)), response_format={ "type": "json_schema", "json_schema": JSONSchema( ## TODO: You can change this depending your task name="ThrillerNovelOutlineSchema", description="A schema for outlining a new novel", schema=task_output_schema, strict=False, ) } ) task_result = json.loads(task_completion.choices[0].message.content)

Aquí está el esquema de nuestro próximo thriller de espías:

 { "storyArc": "Cinderella", "synopsis": "In 'The Clockwork Veil', Ethan Cross, a savvy counter-intelligence operative, is thrust into a high-pressure scenario where leaked documents threaten national integrity. As he races against time to unmask a mole within the agency, Ethan confronts his personal fears and the boundaries of the meticulous strategies he's known for. This slow-burn thriller follows Ethan's transformation in a world where every second could spell disaster, culminating in a showdown with Lena Grey\u2014a former ally who has turned the clockwork of espionage into her personal vendetta. Through grit and cunning, Ethan must adapt his methods, realizing that in the world of espionage, the most powerful weapon is a well-timed intuition.", "characters": [ { "name": "Ethan Cross", "role": "Protagonist", "description": "A meticulous and resourceful counter-intelligence operative known for his analytical mind and calm demeanor under pressure.", "arc": "Ethan transforms from a methodical planner to a decisive action-taker as he confronts his personal fears and realizes the importance of instinct." }, { "name": "Lena Grey", "role": "Antagonist", "description": "A brilliant but disillusioned former operative now turned mole, seeking vengeance against the agency she believes wronged her.", "arc": "Lena starts with a single-minded focus on revenge but gradually becomes conflicted as old loyalties resurface." }, { "name": "Dr. Julia Ward", "role": "Supporting Character", "description": "An astute psychologist who helps Ethan manage the stress of his demanding role and assists in profiling Lena's psychological state.", "arc": "Julia grows from a secondary advisory role to a key player in helping Ethan unearth Lena's motivations." }, { "name": "Michael Garner", "role": "Supporting Character", "description": "Ethan's trusted field partner and an expert in electronic surveillance, providing vital technical support.", "arc": "Michael's experience is tested as he learns to adapt to unpredictable situations, becoming more versatile in his approach." } ], "plotPoints": { "exposition": "Ethan Cross is tasked with investigating a series of leaked documents that could compromise national security. The leaks point to an insider within the agency.", "risingAction": "As Ethan dives deeper, he uncovers a trail leading to Lena Grey, a former colleague presumed dead. Evidence mounts as Ethan closes in, forcing him to question his long-standing methodologies.", "climax": "Ethan finally confronts Lena, who has rigged a trap to destroy critical evidence. In a tense standoff, Ethan must choose between following protocol or taking a risk to stop her.", "fallingAction": "With quick thinking and a new reliance on intuition, Ethan manages to disarm the trap. Lena, deflated, questions her own motives as old memories of camaraderie surface.", "resolution": "Lena is apprehended, the mole hunt ends, and Ethan reflects on his journey, acknowledging the balance between calculated strategy and spontaneity." }, "centralConflict": "Ethan Cross must identify and capture a mole within the agency who is leaking classified information, while dealing with his own rigid attachment to protocol in a dynamically evolving threat landscape.", "setting": "The story unfolds across various global locations including the bustling intelligence hub of Langley, a remote cabin in the Swiss Alps, and the teeming streets of Berlin, lending an authentic and international scope to the narrative.", "majorTurningPoints": [ "Ethan discovers the identity of the mole as his former colleague Lena Grey.", "Lena executes a series of diversions leading to a crisis within the agency.", "Ethan's adherence to protocol nearly costs him a critical breakthrough.", "Ethan's confrontation with Lena culminates in an uncharacteristic display of intuition that saves the mission." ], "characterArcs": { "protagonistArc": "Ethan evolves from strictly adhering to procedures to embracing a balance between strategy and instinctive decision-making, essential in high-stakes situations.", "antagonistArc": "Lena's journey from spite-fueled revenge to questioning her own motivations reflects a shift from isolation to an internal struggle with her past loyalties.", "supportingCharactersArcs": [ "Julia grows from providing psychological insights to playing an active role in strategizing the final approach to Lena.", "Michael transitions from a technical support role to becoming a crucial element in executing Ethan\u2019s plans, emphasizing adaptability." ] } } d

Ahora podemos reutilizar este esquema para generar múltiples soluciones en una secuencia con fuertes garantías acerca de los campos que contiene la salida.

Conclusión

La metaindicación estructurada le permite definir la estructura sobre la marcha, lo que hace que los resultados de LLM sean más confiables para los procesos posteriores. Esté atento a la próxima publicación, donde exploraremos la combinación de metaindicación estructurada con otras técnicas.