Gen-AI chatbots and test-automation of generated responses by mdati

Description

Start experimenting the generative SUSE-AI chat bot, asking questions on different areas of knowledge or science and possibly analyze the quality of the LLM model response, specific and comparative, checking the answers provided by different LLM models to a same query, using proper quality metrics or tools or methodologies.

Try to define basic guidelines and requirements for quality test automation of AI-generated responses.

First approach of investigation can be based on manual testing: methodologies, findings and data can be useful then to organize valid automated testing.

Goals

• Identify criteria and measuring scales for assessment of a text content.
• Define quality of an answer/text based on defined criteria .
• Identify some knowledge sectors and a proper list of problems/questions per sector.
• Manually run query session and apply evaluation criteria to answers.
• Draft requirements for test automation of AI answers.

Resources

• Announcement of SUSE-AI for Hack Week in Slack
• Openplatform and related 3 LLM models gemma:2b, llama3.1:8b, qwen2.5-coder:3b.

Notes

• Foundation models (FMs):
  are large deep learning neural networks, trained on massive datasets, that have changed the way data scientists approach machine learning (ML). Rather than develop artificial intelligence (AI) from scratch, data scientists use a foundation model as a starting point to develop ML models that power new applications more quickly and cost-effectively.

• Large language models (LLMs):
  are a category of foundation models pre-trained on immense amounts of data acquiring abilities by learning statistical relationships from vast amounts of text during a self- and semi-supervised training process, making them capable of understanding and generating natural language and other types of content , to perform a wide range of tasks.
  LLMs can be used for generative AI (artificial intelligence) to produce content based on input prompts in human language.

Validation of a AI-generated answer is not an easy task to perform, as manually as automated.
An LLM answer text shall contain a given level of informations: correcness, completeness, reasoning description etc.
We shall rely in properly applicable and measurable criteria of validation to get an assessment in a limited amount of time and resources.

---

ghostwrAIter - a local AI assisted tool for helping with support cases by paolodepa

Description

This project is meant to fight the loneliness of the support team members, providing them an AI assistant (hopefully) capable of scraping supportconfigs in a RAG fashion, trying to answer specific questions.

Goals

• Setup an Ollama backend, spinning one (or more??) code-focused LLMs selected by license, performance and quality of the results between:
  • deepseek-coder-v2
  • dolphin-mistral
  • starcoder2
  • (...others??)
    
• Setup a Web UI for it, choosing an easily extensible and customizable option between:
  • openwebui
  • localAI
  • lobachat
  • (...others??)
    
• Extend the solution in order to be able to:
  • Add ZIU/Concord shared folders to its RAG context
  • Add BZ cases, splitted in comments to its RAG context
    • A plus would be to login using the IDP portal to ghostwrAIter itself and use the same credentials to query BZ
  • Add specific packages picking them from IBS repos
    • A plus would be to login using the IDP portal to ghostwrAIter itself and use the same credentials to query IBS
    • A plus would be to desume the packages of interest and the right channel and version to be picked from the added BZ cases

---

SUSE AI Meets the Game Board by moio

Use tabletopgames.ai’s open source TAG and PyTAG frameworks to apply Statistical Forward Planning and Deep Reinforcement Learning to two board games of our own design. On an all-green, all-open source, all-AWS stack!

---

Results: Infrastructure Achievements

We successfully built and automated a containerized stack to support our AI experiments. This included:

• a Fully-Automated, One-Command, GPU-accelerated Kubernetes setup: we created an OpenTofu based script, tofu-tag, to deploy SUSE's RKE2 Kubernetes running on CUDA-enabled nodes in AWS, powered by openSUSE with GPU drivers and gpu-operator
• Containerization of the TAG and PyTAG frameworks: TAG (Tabletop AI Games) and PyTAG were patched for seamless deployment in containerized environments. We automated the container image creation process with GitHub Actions. Our forks (PRs upstream upcoming):
  • https://github.com/moio/TabletopGames/tree/hackweek24
  • https://github.com/moio/PyTAG/tree/hackweek24

./deploy.sh and voilà - Kubernetes running PyTAG (k9s, above) with GPU acceleration (nvtop, below)

Results: Game Design Insights

Our project focused on modeling and analyzing two card games of our own design within the TAG framework:

• Game Modeling: We implemented models for Dario's "Bamboo" and Silvio's "Totoro" and "R3" games, enabling AI agents to play thousands of games ...in minutes!
• AI-driven optimization: By analyzing statistical data on moves, strategies, and outcomes, we iteratively tweaked the game mechanics and rules to achieve better balance and player engagement.
• Advanced analytics: Leveraging AI agents with Monte Carlo Tree Search (MCTS) and random action selection, we compared performance metrics to identify optimal strategies and uncover opportunities for game refinement .
  • more about Bamboo on Dario's site
  • more about R3 on Silvio's site (italian, translation coming)
  • more about Totoro on Silvio's site

A family picture of our card games in progress. From the top: Bamboo, Totoro, R3

Results: Learning, Collaboration, and Innovation

Beyond technical accomplishments, the project showcased innovative approaches to coding, learning, and teamwork:

• "Trio programming" with AI assistance: Our "trio programming" approach—two developers and GitHub Copilot—was a standout success, especially in handling slightly-repetitive but not-quite-exactly-copypaste tasks. Java as a language tends to be verbose and we found it to be fitting particularly well.
• AI tools for reporting and documentation: We extensively used AI chatbots to streamline writing and reporting. (Including writing this report! ...but this note was added manually during edit!)
• GPU compute expertise: Overcoming challenges with CUDA drivers and cloud infrastructure deepened our understanding of GPU-accelerated workloads in the open-source ecosystem.
• Game design as a learning platform: By blending AI techniques with creative game design, we learned not only about AI strategies but also about making games fun, engaging, and balanced.

Last but not least we had a lot of fun! ...and this was definitely not a chatbot generated line!

The Context: AI + Board Games

---

Make more sense of openQA test results using AI by livdywan

Description

AI has the potential to help with something many of us spend a lot of time doing which is making sense of openQA logs when a job fails.

User Story

Allison Average has a puzzled look on their face while staring at log files that seem to make little sense. Is this a known issue, something completely new or maybe related to infrastructure changes?

Goals

• Leverage a chat interface to help Allison
• Create a model from scratch based on data from openQA
• Proof of concept for automated analysis of openQA test results

Bonus

• Use AI to suggest solutions to merge conflicts
  • This would need a merge conflict editor that can suggest solving the conflict
• Use image recognition for needles

Resources

• TensorFlow
• open-webui
• openQA instance of openSUSE
• openQA documentation

Timeline

Day 1

• Conversing with open-webui to teach me how to create a model based on openQA test results
  • Asking for example code using TensorFlow in Python
  • Discussing log files to explore what to analyze
  • Drafting a new project called Testimony (based on Implementing a containerized Python action) - the project name was also suggested by the assistant

Day 2

• Using NotebookLLM (Gemini) to produce conversational versions of blog posts
  • Writing openQA tests in Python
  • Developing in distrobox containers
• Researching the possibility of creating a project logo with AI
  • Asking open-webui, persons with prior experience and conducting a web search for advice

Highlights

• I briefly tested compared models to see if they would make me more productive. Between llama, gemma and mistral there was no amazing difference in the results for my case.
• Convincing the chat interface to produce code specific to my use case required very explicit instructions.
• Asking for advice on how to use open-webui itself better was frustratingly unfruitful both in trivial and more advanced regards.
• Documentation on source materials used by LLM's and tools for this purpose seems virtually non-existent - specifically if a logo can be generated based on particular licenses

Outcomes

• Chat interface-supported development is providing good starting points and open-webui being open source is more flexible than Gemini. Although currently some fancy features such as grounding and generated podcasts are missing.
• Allison still has to be very experienced with openQA to use a chat interface for test review. Publicly available system prompts would make that easier, though.

---

Use local/private LLM for semantic knowledge search by digitaltomm

Description

Use a local LLM, based on SUSE AI (ollama, openwebui) to power geeko search (public instance: https://geeko.port0.org/).

Goals

Build a SUSE internal instance of https://geeko.port0.org/ that can operate on internal resources, crawling confluence.suse.com, gitlab.suse.de, etc.

Resources

Repo: https://github.com/digitaltom/semantic-knowledge-search

Public instance: https://geeko.port0.org/

Results

Internal instance:

I have an internal test instance running which has indexed a couple of internal wiki pages from the SCC team. It's using the ollama (llama3.1:8b) backend of suse-ai.openplatform.suse.com to create embedding vectors for indexed resources and to create a chat response. The semantic search for documents is done with a vector search inside of sqlite, using sqlite-vec.

---