Prototype · Ingest validated locally

LLM ForgeBot

An AI workflow prototype with a locally validated document-ingest step

Project Overview

What It Is

A React and Vite builder with Python ingestion and experimental AI tooling. The recorded browser check covers text upload, extraction, and chunk feedback.

Why It Was Built

Explore a shared interface for document preparation and model configuration before separate inference or training runs.

Problem Addressed

Uploaded text and extraction failures need to be visible before documents enter an AI workflow.

Prototype Scope

📄

Document Ingest

A local HTTP service extracts and chunks documents. A text-file upload was recorded in desktop Chrome.

🔍

Upload Feedback

The builder displays the extracted excerpt, chunk count, and upload status.

🎛️

Builder Interface

React and TypeScript screens organize configuration and document preparation into steps.

💬

AI Integration Code

LLM, voice, and image integrations require provider setup and separate runtime validation.

📊

Analytics Scaffold

The backend includes analysis and SQLite demo code, without a verified business-impact claim.

🤖

Training Scripts

SFT and LoRA scripts are present. The ingest check did not train a model.

Architecture Snapshot

Frontend

React and TypeScript served by Vite

ReactTypeScriptViteTailwind CSS

Ingest Service

A local standard-library HTTP server calls the extraction module

PythonThreadingHTTPServerExtractionChunk feedback

Experimental AI Code

Model integrations and training scripts require separate validation

LLM integrationsSFT scriptsLoRA scriptsProvider configuration

Prototype Storage

The broader backend includes a SQLite demo database and local artifacts

SQLiteSQLAlchemyLocal filesDemo data

Verified API Scope

The recorded browser flow uses ingestion endpoints without training

GET /healthPOST /ingestLocal CORS scopeUpload size limit

Tech Stack

Frontend

• React

• TypeScript

• Vite

• Tailwind CSS

Ingestion

• Python

• HTTP server

• Text extraction

• Chunking

Experimental Backend

• FastAPI source

• SQLite

• SQLAlchemy

Training Code

• PyTorch

• SFT

• LoRA

Validation

• Local ingestion

• Browser upload feedback

• Component tests

Challenges & Learnings

Separating Ingest from Training

Problem

A successful upload does not establish a completed training run.

Solution

The local ingest service returns extraction results and explicitly does not train.

Learning

Validation evidence should name the exact workflow exercised.

Inspectable Upload Results

Problem

Users need to see what text was recovered from their files.

Solution

Show the extracted excerpt, status, and chunk count after the ingestion call.

Learning

Visible feedback supports review before downstream processing.

Evidence Boundaries

Problem

Integration scripts alone do not establish an operational AI service.

Solution

Describe the browser ingest check separately from experimental model integrations.

Learning

User trials and model execution require separate evidence.

Target Users

Document Preparation

Inspect extracted text and chunk feedback locally

AI Workflow Experiments

Explore builder and integration code after runtime configuration

Training Research

Review training scripts separately from the validated ingestion path

Scope & Validation

Verified Scope

✓Repository evidence records a desktop-browser text ingest check
✓The local endpoint returns extracted content and chunk counts
✓The repository records ingestion and upload-motion checks
✓Model training and independent-user trials remain unverified

Engineering Focus

→Document ingestion feedback
→React and TypeScript interfaces
→Separating UI and model validation
→Documenting prototype boundaries

Validation Boundary

The recorded validation covers document ingestion in a local desktop browser. It does not establish model training, production deployment, measured cost savings, or operational performance gains.