Learning Ollama

Just one of the things I'm learning. https://github.com/hchiam/learning

Ollama makes it easy to run LLMs offline/locally/privately on your computer.

Setup

1. ollama.com website download app to install ollama command
2. ollama run llama2 automatically downloads llama2 model if needed, and lets you talk with the model offline directly in the terminal, with the ollama app running in the background
   • (otherwise to just use another terminal window to run ollama in the background: ollama serve)

Extra notes

• in .bash_profile: alias llm='ollama run llama2'
• to read a local file: ollama run llama2 "Summarize this file: $(cat scratchpad.md)"
• or with the custom alias: llm "Summarize this file: $(cat scratchpad.md)"
• or run ollama with python code: https://github.com/hchiam/local-llms-analyse-finance/blob/main/categorize_expenses.ipynb
• but overall, it seems:
  • running llama2 on ollama has limited speed (unless your computer has GPUs that can Ollama can use for GPU acceleration)
  • running llama2 on ollama has limited prompt/input size, which would require chopping up the input

Further resources

• https://youtu.be/WxYC9-hBM_g?feature=shared (around 18:08)
  • (aside: consider privategpt too:)
    • https://dev.to/docteurrs/installing-privategpt-on-wsl-with-gpu-support-1m2a
    • and https://docs.privategpt.dev/overview/welcome/introduction
• https://www.youtube.com/watch?v=h_GTxRFYETY
• https://ollama.com
• to customize with a Modelfile (temperature and system message/prompt): https://github.com/ollama/ollama?tab=readme-ov-file#customize-a-prompt

Example uses

• use ollama to set up privateGPT
• use ollama to set up Gemma2 to run with local RAG using LangChain with samwit's code to run (or my fork of that repo):

  • pip install langchain
    # langchain_community might need to installed with pip3 install langchain-community langchain-core

  • ollama run gemma2 # 5.4 GB when i tried
    ollama pull nomic-embed-text # 274 MB
    ollama list

  • python indexer.py

  • python ollama_gemma2_rag_debugging.py

    or

    python ollama_gemma2_rag_simple.py