prompt-identifiers-ai-sdk

Vercel AI SDK middleware for automatic ID encoding/decoding in LLM prompts. Reduces token usage by up to 90% for UUIDs and ULIDs.

Installation

npm install prompt-identifiers-ai-sdk prompt-identifiers ai

Quick Start

import { openai } from "@ai-sdk/openai";
import { wrapLanguageModel, generateText } from "ai";
import { promptIdentifiersMiddleware } from "prompt-identifiers-ai-sdk";

// Wrap your model with the middleware
const model = wrapLanguageModel({
  model: openai("gpt-4o"),
  middleware: promptIdentifiersMiddleware({
    config: { inputFormat: "UUID", outputFormat: "SafeNumeric" },
  }),
});

// Use normally - IDs are automatically encoded/decoded
const result = await generateText({
  model,
  prompt: "Summarize activity for user 123e4567-e89b-42d3-a456-426655440000",
});

// The LLM sees: "Summarize activity for user ~000~"
// You receive the response with original UUIDs restored

How It Works

1. Before LLM call: The middleware encodes IDs in your prompt messages

   • 123e4567-e89b-42d3-a456-426655440000 → ~000~

2. After LLM call: The middleware decodes IDs in the response

   • ~000~ → 123e4567-e89b-42d3-a456-426655440000

This is completely transparent - you work with real IDs, the LLM works with compact placeholders.

Configuration

promptIdentifiersMiddleware({
  // Required: encoding configuration
  config: {
    inputFormat: "UUID", // or 'ULID' or custom RegExp
    outputFormat: "SafeNumeric", // or 'Numeric', 'IdToken', { template: '...' }
  },

  // Inject a format-preservation instruction into the system message (default: true)
  // Prevents LLMs from stripping delimiter characters in tool call arguments
  injectInstruction: true,

  // Custom instruction text ({format} is replaced with example token like ~000~)
  customInstruction: "IDs use {format} notation. Preserve them exactly.",

  // Optional: enable debug mode for detailed diagnostics
  debug: true,

  // Optional: callbacks for logging/debugging
  onEncode: (result) => {
    console.log("Mapping:", result.mapping);
    if (result.debugData) {
      console.log(`Encoded ${result.debugData.encodedCount} IDs in ${result.debugData.durationMs}ms`);
    }
  },
  onDecode: (result) => {
    console.log("Decoded output:", result.output);
    // Warnings detect anomalies like stripped delimiters
    if (result.warnings?.length) {
      console.warn("Decode warnings:", result.warnings);
    }
    if (result.debugData) {
      console.log(`Decoded ${result.debugData.decodedCount} placeholders in ${result.debugData.durationMs}ms`);
    }
  },
});

Input Formats

| Format | Description | Example | | -------- | ---------------------- | -------------------------------------- | | 'UUID' | RFC 4122 UUIDs | 123e4567-e89b-42d3-a456-426655440000 | | 'ULID' | Crockford Base32 ULIDs | 01ARZ3NDEKTSV4RRFFQ69G5FAV | | RegExp | Custom pattern | /user-\d{6}/gi |

Output Formats

| Format | Description | Example | | --------------------- | ------------------------------------------------- | ------------------------------------- | | 'SafeNumeric' | Collision-safe with tildes (recommended) | ~000~, ~001~ | | 'Numeric' | Simple numeric with smart triplet expansion | 000, 001 | | 'IdToken' | Base62 encoding | 0, A, z, 10 | | { template: '...' } | Custom template | { template: '[ID:{i}]' } → [ID:0] |

Streaming Support

Important: For streaming responses, use a delimited format like SafeNumeric or a custom template with delimiters. Non-delimited formats (Numeric, IdToken) cannot reliably handle placeholders split across chunks.

The middleware buffers incomplete placeholders when using delimited formats like SafeNumeric:

import { streamText } from "ai";

const result = await streamText({
  model, // wrapped model
  prompt: "Analyze user 123e4567-e89b-42d3-a456-426655440000",
});

for await (const chunk of result.textStream) {
  // IDs are decoded in real-time as the stream arrives
  console.log(chunk);
}

Works With Any Provider

The middleware works with any AI SDK provider:

import { anthropic } from "@ai-sdk/anthropic";
import { google } from "@ai-sdk/google";
import { mistral } from "@ai-sdk/mistral";

// Anthropic
const claudeModel = wrapLanguageModel({
  model: anthropic("claude-3-5-sonnet-20241022"),
  middleware: promptIdentifiersMiddleware({ config }),
});

// Google
const geminiModel = wrapLanguageModel({
  model: google("gemini-1.5-pro"),
  middleware: promptIdentifiersMiddleware({ config }),
});

Multiple IDs

Duplicate IDs are automatically deduplicated:

const prompt = `
  User 123e4567-e89b-42d3-a456-426655440000 sent a message.
  Check if 123e4567-e89b-42d3-a456-426655440000 is online.
  User 987fcdeb-51a2-43f7-8d9c-0123456789ab received it.
`;

// Encoded as:
// User ~000~ sent a message.
// Check if ~000~ is online.
// User ~001~ received it.

License

MIT