VENI-AI

English

Tiếng Việt

English

Tiếng Việt

Appearance

Technical Specification: AI Assistant ​

This document details the technical architecture, RAG pipeline, and infrastructure for the AI Assistant module.

v1.0 — ApprovedPlatform: VENI-AI ShellStack: Bun/Ignis/Hono

1. High-Level Architecture ​

The AI Assistant is an RAG-based (Retrieval-Augmented Generation) system. It utilizes a vector database to perform semantic search over organization-scoped data before generating responses via a Large Language Model (LLM).

Component Diagram ​

2. Technology Stack ​

Backend (API) ​

  • Runtime: Bun
  • Framework: Ignis Framework
  • Router: Hono
  • Vector DB: PostgreSQL with pgvector extension.
  • LLM Engine: OpenAI (GPT-4o / GPT-4o-mini).
  • Embeddings: text-embedding-3-small (1536 dimensions).

Frontend (UI) ​

  • Library: React 18
  • Build Tool: Vite
  • Styling: Tailwind CSS v4 + ARDOR UI Kit
  • Streaming: Server-Sent Events (SSE) for real-time response rendering.

3. RAG Implementation Details ​

3.1 Indexing Pipeline (Knowledge Base) ​

  1. Extraction: Raw text is pulled from Drive (via S3 streaming) or Document (via SQL).
  2. Chunking: Text is split into overlapping chunks (e.g., 500 tokens with 50-token overlap).
  3. Embedding: Chunks are sent to OpenAI's embedding API.
  4. Persistence: Vectors and text chunks are stored in the kb_chunks table, isolated by orgId.

3.2 Retrieval Pipeline (Q&A) ​

  1. Query Embedding: The user's question is embedded into a 1536-dim vector.
  2. Vector Search: A cosine similarity search is performed using pgvector's <=> operator.
  3. Context Assembly: The top K most relevant chunks are retrieved and formatted into the LLM system prompt.
  4. Generation: The LLM generates the answer, following strict instructions to cite sources only from the provided context.

4. Security & Privacy ​

4.1 Data Isolation ​

Every knowledge base source and its corresponding chunks are tagged with an organizationId. The vector search queries are strictly scoped to the user's organization.

4.2 Training Opt-out ​

VENI-AI uses Enterprise API agreements with LLM providers ensuring that data sent for completion or embedding is never used for training public models.

4.3 Citation Integrity ​

The system implements a "Contextual Highlight" feature that fetches the source chunk directly from the database when a citation is clicked, ensuring the user sees exactly what the AI read.

Related Links

© 2025-2026 VENIZIA AI