ai-ml/editorial-guide-ramalama
2
4
Fork 2
Code Issues Pull requests Releases 1 Activity Actions
1 release 1 tag
  • v1.0.0 07f5414890

    v1.0.0 - Fedora Editorial Guide RAG Assistant Stable

    gtfrans2re released this 2026-07-13 17:19:01 +00:00 | 4 commits to main since this release

    v1.0.0 - Fedora Editorial Guide RAG Assistant

    An AI editorial assistant that reviews draft articles against the Fedora Community Blog and Fedora Magazine writing guidelines, using a local Retrieval-Augmented Generation (RAG) pipeline built on RamaLama.

    What's included

    Pipeline

    • RamaLama-based RAG pipeline (fetch -> stage -> build) over the Fedora Community Blog and Fedora Magazine writing guidelines and past articles
    • Docling-based document conversion and chunking (chunk-size 100)
    • Pre-built vector-store images published to Quay

    Interfaces

    • Streamlit web app (app.py) - the primary entry point for end users
    • Command-line usage via ramalama run --rag
    • Jupyter notebook for exploration

    Models

    • Gemma 4 E4B (hf://bartowski/google_gemma-4-E4B-it-GGUF)
    • Granite 4 Micro 3B (hf://ibm-granite/granite-4.0-micro-GGUF)

    Evaluation

    • Local evaluation harness (eval/) scoring the pipeline with both RAGAS and DeepEval using a local judge model - no external API calls
    • Guideline-grounded evaluation set and side-by-side comparison tooling

    Documentation

    • docs/QUICKSTART.md - run the pre-built images, no pipeline setup needed
    • docs/GUIDE.md - full pipeline walkthrough
    • docs/OPENSHIFT_CRC.md / docs/COMMUNISHIFT.md - deployment guidance
    • docs/INFRA_APPRENTICE.md - Fedora infrastructure apprentice runbook
    • docs/CONTRIBUTING.md - contribution workflow
    • AGENTS.md - guidance for AI coding agents

    Container images

    All images are published to the Fedora organization on Quay:

    Tag Corpus
    quay.io/fedora/editorial-guide-ramalama:commblog Fedora Community Blog
    quay.io/fedora/editorial-guide-ramalama:magazine Fedora Magazine
    quay.io/fedora/editorial-guide-ramalama:editorial Combined

    Quick start:

    podman pull quay.io/fedora/editorial-guide-ramalama:commblog
ramalama run --rag quay.io/fedora/editorial-guide-ramalama:commblog \
  hf://ibm-granite/granite-4.0-micro-GGUF

    Known limitations

    • Retrieved contexts are not yet captured in the evaluation harness, so only answer relevancy is scored. Faithfulness and context precision/recall require context capture (planned).
    • Memory-intensive. Running the pipeline and a judge model concurrently can exhaust RAM on a typical laptop. Use the lighter Granite model and run phases separately, or use dedicated hardware.
    • RAGAS requires Python 3.12 (incompatible with Python 3.14's asyncio changes) plus a separate local embeddings model. DeepEval runs without these constraints.
    • Not deployed to Communishift. Deployment guidance is documented, but a namespace request is still pending.

    Credits

    Developed by Francois Gonothi Toure (@gtfrans2re) as an Outreachy 2026 internship mini-project with the Fedora Project.

    Mentors: @jflory7, @cybette, @dominikkawka

    Thanks to the Fedora AI/ML SIG, the Fedora Community Blog and Fedora Magazine editors, and the RamaLama project.

    Assisted-by: Claude Opus 4.8

    Downloads