Giselle AI Agent Studio

Giselle is an AI agent studio designed for agentic workflows and human-AI collaboration. Key features include:

• GitHub AI Operations: Automates issues, PRs, and deployments.
• Visual Agent Builder: Drag-and-drop interface for creating and modifying agents.
• Multi-Model Composition: Uses multiple models (GPT, Claude, Gemini, etc.) to select the best model for specific tasks.
• Knowledge Store: Centralized access to code and data, with GitHub vector store integration.
• Team Collaboration (In Development): Shared configurations and contextual awareness for teams.
• Template Hub (In Development): Community-contributed agent templates.

Use @giselles-ai/pseudo-tiktoken to get approximate token counts for text. This is a lightweight, TypeScript-only implementation designed to estimate counts for models like gpt-4o-mini without the overhead of WASM or the actual tiktoken library.

Important: This is a pseudo-implementation for estimation purposes (e.g., UI displays, safe context length management, or input truncation). It is not a complete replacement for tiktoken and may have a relative error of approximately ±25%. It is intentionally designed to slightly overestimate counts to ensure safety in context length management.

import { countTokens } from "@giselles-ai/pseudo-tiktoken";

const text = "Hello, world! This is a test.";
const tokenCount = countTokens(text);
console.log(`Token count: ${tokenCount}`);

Giselle uses a Vault system to securely manage sensitive information like Personal Access Tokens (PATs) or LLM API keys. The Vault provides encrypt and decrypt capabilities and is integrated into the GiselleEngine via a driver-based architecture.

To use a Vault, you must provide a Vault implementation (created via createVault using a VaultDriver) when initializing the engine.

Note: Decryption is a server-side only operation. The client-side API only exposes encryption methods.

import { createVault, VaultDriver } from '@giselles-ai/giselle-engine';

// Define or import a driver
const myDriver: VaultDriver = { /* implementation */ };

// Create the vault instance
const vault = createVault(myDriver);

// Pass to the engine
const engine = NextGiselleEngine({
  vault,
  // other config...
});

All public APIs in this package accept an optional signal property in their options object. This allows you to pass an AbortSignal to cancel ongoing text extraction or image rendering, which is critical for managing resource-intensive ingestion flows.

const abortController = new AbortController();
setTimeout(() => abortController.abort("timeout"), 5_000);

await extractPdfText(data, {
  signal: abortController.signal,
  pdfiumWasmBinary,
});

When running an app via the SDK, you can provide file inputs in two ways:

1. fileId: Use an ID obtained from the upload API.
2. Inline base64: Provide a base64 encoded file payload directly (subject to size limits).

The tokenizer follows a three-step process to estimate tokens:

1. Whitespace normalization: Converts newlines, tabs, and consecutive spaces into single spaces.
2. Lexical splitting:
   • Whitespace: Ignored and not counted as tokens.
   • Punctuation: Each punctuation character is treated as 1 token.
   • English words: Alphanumeric words (including apostrophes) are split into subwords.
   • Other characters: Characters like emojis or CJK are treated as 1 token each.
3. Subword splitting: Long English words are split using common suffix patterns (e.g., tokenization → token + ization) or fixed-size chunks.

The project separates system-level Supabase configurations from application-level database schemas using two distinct directories:

• migrations/system/: Contains .sql files for Supabase system schema migrations (e.g., auth, storage). These are managed manually.
• migrations/schema/: Contains application schema migrations, which are managed by Drizzle.

migrations/
  system/             # Supabase system schema migrations
    *.sql            # SQL migration files
  schema/            # Application schema migrations (managed by Drizzle)

When styling components in Giselle, follow these migration and maintenance rules:

1. Avoid raw colors: Do not use hex codes or raw color primitives directly in your styles. Always prefer semantic tokens and utilities.
2. Use the v3 bridge: Continue using the v3 bridge in aliases.css.
3. Minimize aliases: Only add new aliases in aliases.css when absolutely necessary during the migration process.