grapes-ai

grapes-ai

A post-processing and transformation plugin framework for AI outputs and API responses.

grapes-ai is a lightweight, chainable processing pipeline designed to validate, repair, and transform data. It is built for engineering teams who need deterministic guarantees when working with probabilistic AI outputs.

What is grapes-ai?

It is a framework for building "Grapes"—single-purpose logical units that sit between your raw data source (LLM, API, database) and your application logic.

Each Grape:

• Validates: Checks if data meets specific criteria (e.g., schema, aspect ratio).
• Enforces: Mutates data to satisfy constraints (e.g., resizing images, casting types).
• Transforms: Converts data into your required format.
• Fails: Provides explicit failure signals when constraints cannot be met.

What grapes-ai is NOT

• NOT an Agent Framework: It does not plan, reason, or operate autonomously. It follows a strict linear pipeline.
• NOT a "Fact Checker": It cannot verify if an AI's statement is true. It can only verify if the output matches a specific structure or format.
• NOT a Model Trainer: It operates purely on the output side.
• NOT Magic: It relies on defined engineering rules/schema, not "AI reasoning" to fix problems.

Core Concepts

1. Grape

A small, functional module.

interface Grape<Input, Output> {
  process(input: Input): Promise<Result<Output, GrapeError>>;
}

API Reference

The Result Type

We strictly avoid throwing Errors for control flow. All Grapes must return a Result.

type Result<T, E = Error> =
  | { success: true; data: T }       // Operation succeeded
  | { success: false; error: E };    // Operation failed explicitly

This discriminative union forces you to handle both cases. accessing .data is only allowed if success is true.

Error Propagation

• Fatal: If a Grape returns success: false, the Pipeline halts immediately.
• Recoverable: If you need to recover, you must handle the logic outside the pipeline or use a specific "RetryGrape" (planned) that wraps another Grape.

2. Pipeline

A linear chain of Grapes. Data flows through sequentially. If any Grape fails, the pipeline halts (or falls back, depending on configuration).

3. Failure

Errors are first-class citizens. A pipeline returning a failure is a valid state, allowing your application to trigger fallbacks or retries deterministically.

Installation

npm install grapes-ai

Usage Examples

1. Text Post-Processing (JSON Schema Enforcement)

Ensure an LLM response is valid JSON and adheres to a specific schema.

import { Pipeline, JsonParserGrape, SchemaValidatorGrape } from 'grapes-ai';
import { z } from 'zod';

const userSchema = z.object({
  id: z.number(),
  username: z.string(),
});

const pipeline = new Pipeline()
  .add(new JsonParserGrape())            // String -> Object
  .add(new SchemaValidatorGrape(userSchema)); // Object -> Typed User

async function handleLlmOutput(rawString: string) {
  const result = await pipeline.run(rawString);

  if (result.success) {
    console.log("Valid User:", result.data);
  } else {
    console.error("Validation Failed:", result.error);
    // Explicitly handle failure (e.g., retry LLM call)
  }
}

2. API Transformation

Normalize "messy" API data into a clean internal type.

import { Pipeline, RenameKeyGrape, TypeCastGrape } from 'grapes-ai';

const apiPipeline = new Pipeline()
  .add(new RenameKeyGrape('user_id', 'userId'))
  .add(new TypeCastGrape('userId', 'number'))
  .add(new DefaultValueGrape('isActive', false));

const cleanData = await apiPipeline.run(externalApiData);

3. Image Post-Processing

Validate and standardize image assets before storage.

import { Pipeline, AspectRatioGrape, MaxSizeGrape } from 'grapes-ai';

const imagePipeline = new Pipeline()
  .add(new AspectRatioGrape(16 / 9, { tolerance: 0.05 })) // Fails if not close to 16:9
  .add(new MaxSizeGrape(1920, 1080)); // Downscales if larger

const processedImage = await imagePipeline.run(uploadedBuffer);

Design Philosophy

1. Deterministic First, AI Second

We trust code more than probabilities.

• Why? LLMs are non-deterministic. If you ask an LLM to "fix this JSON," it might fix it 99 times and hallucinate a new key the 100th time.
• Approach: Always use regex, strict schema validation (Zod), or algorithmic repair first. Only use checking models as a last resort "Heal Grape."

2. Explicit Failures > Silent Guesses

A pipeline should crash loudly rather than pass bad data.

• Failures Propagate: A failure in the first Grape stops the entire chain immediately.
• Signal: We return a { success: false, error: Error } Result type. We never throw exceptions for expected validation failures; we return them as values.

3. Isolated & Chainable

Big logic is bad logic.

• Small Grapes: A Grape should do one thing (e.g., "Check Aspect Ratio").
• Composable: If a rule changes, you remove one Grape, not rewrite a monolithic function.

Limitations

• No Semantic Truth: grapes-ai cannot tell you if a generated summary is accurate, only that it is grammatically correct or fits a length constraint.
• Linear Complexity: Pipelines are designed for linear transformations. Complex branching logic logic should be handled in your application code, not creating "spaghetti grapes."
• Misuse of Fallbacks:
  • Anti-Pattern: Using an LLM to "fix" a date format that could be fixed with Date.parse().
  • Rule: Never call an LLM if you can fix it with deterministic logic. Use "Healer Grapes" only when schema validation fails explicitly and no algorithmic fix is possible.

Contributing

We welcome pull requests that align with our deterministic philosophy.

Standards for New Grapes

• Single Responsibility: A Grape must do exactly one thing.
• No Side Effects: Grapes should be pure functions of their input where possible (or idempotent).
• Strict Typing: We enforce noImplicitAny.
  • ❌ process(input: any)
  • ✅ process(input: unknown) or process(input: string)
• Testing:
  • Must include .test.ts.
  • Must test success (happy path) AND failure (edge cases).

Testing & Linting

• Test every Grape: Create a matching .test.ts file in tests/.
• Coverage: run npm test to ensure no regressions.
• Lint: Code must pass tsc with strict settings enabled.

Available Grapes

Module	Grape	Input	Output	Description
Text	JsonParserGrape	string	unknown	Safely parses JSON string
Text	SchemaValidatorGrape	unknown	T	Validates against Zod schema
API	RenameKeyGrape	Record	Record	Renames object keys
API	TypeCastGrape	Record	Record	Casts values to primitives
Image	AspectRatioGrape	ImageMetadata	ImageMetadata	Validation only

Roadmap

• ✅ Core Engine: TypeScript interfaces, Pipeline logic, and explicit Result types.
• ✅ Text Grapes: JSON parsing, Zod schema validation.
• ✅ API Grapes: Key renaming, Type casting, Default values.
• ⚡ Image Processing: (Planned) sharp integration for resizing, aspect ratio, and format conversion.
• 🧠 Healer Grapes: (Planned) Optional LLM-based repair for broken JSON. Philosophy: Must be less trusted than deterministic steps and only run on explicit failure.

Maintained by the Open Source Engineering Community.