LLMBox

Visual Inspection 3.0: The Era of Prompt Training, Zero cloud LLM dependency, 100% local!

Visual Inspection

• 1.0 : The algorithm is made by human
• 2.0 : The weights of CNN are trained
• 3.0 : The prompt of LLM is trained

Tech Stack

• local App: Next.js
• local AI: Ollama
• local DB: Postgres
• local Storage: Minio

How to run

1. Clone the repository
2. Run docker compose up -d
3. Run pnpm install
4. Run pnpm dev

Prerequisites

• Docker
• Ollama

Dashboard

• App: http://localhost:3000/
• Minio: http://localhost:9001/

Development

• DB: https://local.drizzle.studio/
  • Run pnpm run db:studio to open the studio

Entity Relationship Diagram

erDiagram
    projects ||--o{ files : "has"
    projects ||--o{ criterias : "has"
    projects ||--o{ promptTemplates : "has"
    projects ||--o{ inspectionSpecs : "has"
    projects ||--o{ promptEvaluations : "has"

    projects {
        int id PK
        string name
        string description
        timestamp createdAt
        timestamp updatedAt
    }

    files ||--o{ filesToCriterias : "has"
    files ||--o{ evalDetails : "has"
    files {
        int id PK
        int projectId FK
        string fileName
        string originalName
        string mimeType
        int size
        enum uploadType
        enum aiLabel
        int aiPromptId FK
        enum humanLabel
        vector embedding
        timestamp createdAt
    }

    criterias ||--o{ filesToCriterias : "has"
    criterias {
        int id PK
        int projectId FK
        string name
        string description
        timestamp createdAt
    }

    filesToCriterias {
        int fileId FK
        int criteriaId FK
        boolean isFail
        string reason
        timestamp createdAt
    }

    promptTemplates ||--o{ promptEvaluations : "used_in"
    promptTemplates {
        int id PK
        int projectId FK
        string text
        timestamp createdAt
    }

    inspectionSpecs ||--o{ promptEvaluations : "used_in"
    inspectionSpecs {
        int id PK
        int projectId FK
        string text
        timestamp createdAt
    }

    promptEvaluations ||--o{ evalDetails : "has"
    promptEvaluations {
        int id PK
        int projectId FK
        int promptTemplateId FK
        int inspectionSpecId FK
        string finalPrompt
        float score
        enum state
        int duration
        string analysisText
        int numDataset
        timestamp createdAt
    }

    evalDetails {
        int id PK
        int fileId FK
        int promptEvalId FK
        enum llmLabel
        string llmReason
        enum result
        timestamp createdAt
    }

Loading