Skip to content

AI_README.md

Latest commit

 

History

History
109 lines (84 loc) · 3.46 KB

AI_README.md

File metadata and controls

109 lines (84 loc) · 3.46 KB

SAAB Knowledge AI — Local DeepSeek Setup

Run a 100% local AI assistant trained on this SAAB knowledge base using DeepSeek via Ollama.
No cloud, no API keys, no data leaves your machine.

Prerequisites

Requirement Minimum Link
Python 3.10+ https://python.org
Ollama Latest https://ollama.com
RAM 8 GB (16 GB recommended) —
Disk ~6 GB for the 8B model —

Quick Start

One-click setup (Windows)

.\setup.ps1

Manual setup

# 1. Install Python dependencies
pip install -r requirements.txt

# 2. Make sure Ollama is running
ollama serve          # in a separate terminal if not running as a service

# 3. Pull the DeepSeek model
ollama pull deepseek-r1:8b

# 4. Ingest the SAAB docs into the vector store
python ingest.py

# 5. Start chatting
python chat.py

How It Works

Your question
    │
    ▼
┌──────────────────┐     ┌──────────────────┐
│  Embedding Model │────▶│  ChromaDB Vector  │
│  (all-MiniLM-L6) │     │  Store (.vectordb)│
└──────────────────┘     └────────┬─────────┘
                                  │ top-5 relevant chunks
                                  ▼
                         ┌──────────────────┐
                         │  DeepSeek R1 8B  │
                         │  (Ollama local)  │
                         └────────┬─────────┘
                                  │
                                  ▼
                            AI Answer
  1. Ingest (ingest.py): Reads all SAAB .md files, chunks them, embeds them with a local model, and stores vectors in ChromaDB.
  2. Chat (chat.py): Your question is embedded, matched against the knowledge base, and the top matches are sent to DeepSeek as context for answering.

Choosing a Model Size

Edit the MODEL variable in chat.py:

Model VRAM/RAM Quality Speed
deepseek-r1:1.5b ~2 GB Basic Fast
deepseek-r1:8b ~6 GB Good (default) Medium
deepseek-r1:14b ~10 GB Better Slower
deepseek-r1:32b ~20 GB Best Slowest

Pull your chosen model: ollama pull deepseek-r1:14b

Re-ingesting After Data Changes

If you update any .md files, re-run:

python ingest.py

This rebuilds the vector store from scratch.

File Structure

SAAB/
├── 96.md            # SAAB 96 knowledge
├── C900.md          # Classic 900 knowledge
├── NG9-3.md         # New Generation 9-3 knowledge
├── NG9-5.md         # New Generation 9-5 knowledge
├── chat.py          # Interactive chatbot
├── ingest.py        # Document ingestion pipeline
├── setup.ps1        # One-click Windows setup
├── requirements.txt # Python dependencies
├── .vectordb/       # (generated) Vector store
├── Documents/       # Additional reference docs
└── k8s/             # Docker & Kubernetes files
    ├── Dockerfile       # Production container image
    ├── Dockerfile.dev   # Dev sandbox container image
    ├── k8s-deployment.yaml  # K8s standalone chat deploy
    ├── k8s-dev.yaml         # K8s dev sandbox deploy
    └── QuickStart.md        # K8s setup instructions