Open Source Backend for AI Agents

Wipe Coding: Promise, Reality, and the INS Forge Solution

Key Concepts:

• Wipe Coding: The concept of using AI agents to automatically generate and deploy applications from simple prompts.
• Semantic Layer: An abstraction layer that translates between the agent's requests and the complexities of the backend infrastructure.
• INS Forge: An open-source semantic layer designed to simplify backend development for AI agents.
• MCP Server: (Model Call Protocol Server) A server that facilitates communication between the coding agent and the backend infrastructure.
• RAG (Retrieval-Augmented Generation): A technique combining semantic search with LLMs for more informed responses, used in the example application.
• pgvector: A PostgreSQL extension for storing and querying vector embeddings.
• Row Level Policies: Security measures within a database that control access to data based on user roles or other criteria.

The Current State of Wipe Coding: A Disconnect Between Promise and Reality

The video begins by contrasting the marketed promise of wipe coding – prompting an agent to build and deploy a fully functional application – with the current reality. While agents excel at creating functional User Interfaces (UIs), significant challenges arise when attempting to implement a robust backend. The core issue is the complexity of backend infrastructure. Agents frequently “hallucinate” non-existent APIs, generate incorrect configuration files, and ultimately require manual debugging by developers – precisely the task wipe coding aims to eliminate.

The presenter explains that frontend development benefits from well-documented and predictable APIs, offering consistent patterns. However, backend development involves numerous interconnected components (databases, schemas, authentication, etc.), each with its own complexities, making it difficult for agents to navigate successfully. As stated, “These systems try to configure the database… But usually they hallucinate an API that does not exist, writes configuration files that are completely wrong.”

Introducing INS Forge: A Semantic Layer for Backend Development

To address these challenges, the presenter introduces INS Forge, an open-source project designed as a semantic layer between coding agents and the backend. INS Forge, licensed under Apache 2.0, can be self-hosted within a Docker container or utilized via a cloud-based deployment. Its purpose is to abstract away the complexities of backend infrastructure, providing a consistent interface for agents to interact with. The presenter emphasizes that it provides “all the core components that you need for your backend infrastructure.”

Step-by-Step Deployment with INS Forge: A RAG Application Example

The video demonstrates a practical application of INS Forge by deploying a backend for an agentic file search project (RAG). This project combines semantic search (using embeddings) with agentic file search to provide more accurate and contextually relevant results. The backend requires a database to store document metadata, embeddings, and extracted schemas.

The deployment process involves the following steps:

1. Project Creation: A new project named “rag” is created within INS Forge.
2. IDE Integration: The INS IDE extension is installed in Codeex, enabling communication with the INS MCP server.
3. Agent Instruction: Codeex is instructed to deploy the project with a fully functional backend using the INS MCP server, specifying database storage (pgvector/PostgreSQL) and gateway usage for embeddings and LLMs.
4. Automated Backend Setup: INS Forge automatically creates the necessary database tables (corpa, documents, chunks, schema, chunk embeddings), implements row-level security policies, and provides a secure storage for secrets.
5. Visualizer: The INS visualizer provides a clear overview of the entire backend structure, including database schemas and file storage arrangements.
6. Model Gateway Integration: Gemini 3 Flash preview is enabled via the model gateway, along with support for embedding models.
7. UI Interaction: The deployed backend is accessed through a local UI, allowing for document indexing and querying.

The deployment process took 28 minutes using Codeex, highlighting the computational demands of the automated backend setup.

Backend Functionality and Data Flow in the RAG Application

Once deployed, the RAG application functions as follows:

• Document Upload & Processing: Documents are uploaded and stored in the “corpa” table.
• Chunking & Embedding: Documents are divided into chunks, and embeddings are generated and stored in the “chunk embeddings” table.
• Schema Extraction: Schemas are automatically extracted from the documents and stored in the “schema” table.
• Query Processing: User queries are processed using semantic search on the embedding table to identify relevant chunks.
• LLM Integration: The relevant chunks and documents are fed to a Large Language Model (LLM) for agentic file search, providing a comprehensive and contextually aware response.

The presenter demonstrates a query ("What was the purchase price for this acquisition?") and explains how the system retrieves information from the database and utilizes the LLM to provide a detailed answer, referencing specific details from the original document.

Monitoring and Deployment Options

INS Forge provides detailed logs for monitoring the application's performance and identifying potential issues. It also offers a hosted solution for direct application deployment with custom domain integration and environment variable management. The presenter emphasizes the benefit of INS Forge being open-source, allowing users to self-host and maintain complete control over their infrastructure. As stated, “The thing that I like the most about this is that it's open source. So you can actually see what exactly is happening and you don't have to use their platform if you just want to deploy it yourself.”

Conclusion: Bridging the Gap in Wipe Coding

The video concludes that INS Forge represents a significant step towards realizing the promise of wipe coding by addressing the complexities of backend development. By providing a semantic layer that abstracts away the intricacies of infrastructure, INS Forge empowers AI agents to build and deploy fully functional applications more reliably and efficiently. The open-source nature of the project further enhances its value, offering users flexibility and control over their deployments. The key takeaway is that a well-designed backend infrastructure, tailored for agent interaction, is crucial for unlocking the full potential of AI-powered application development.