[RFE] Enable local LLM provider support (Ollama) for development and testing

[anik120]

Is your feature request related to a problem? Please describe.

Currently, Lightspeed Stack only officially supports cloud-based LLM providers (OpenAI, Azure, RHOAI, RHEL AI). This creates significant barriers for developers and users trying to:

• Get started with the project
• Test and develop locally
• Control costs during development
• Work offline or in air-gapped environments
• Learn the system without cloud dependencies

Current Situation

From docs/providers.md:

Provider	Type	Supported in LCS
ollama	remote	❌
meta-reference	inline	❌
sentence-transformers	inline	❌

Developers currently must:

1. Create an OpenAI/Azure account
2. Add payment information
3. Manage API keys and quotas
4. Pay for every test query during development
5. Wait for quota resets when limits are hit

Example: A developer following the getting started guide encounters:

RateLimitError: You exceeded your current quota

This blocks them from testing basic functionality without adding credits.

Describe the solution you'd like

Add official support for Ollama as a local inference provider.

Ollama provides:

• Free, unlimited local inference (no API keys, no quotas)
• Easy installation (brew install ollama / curl -fsSL https://ollama.ai/install.sh | sh)
• Production-quality models (Llama 3, Mistral, Phi, etc.)
• OpenAI-compatible API (minimal integration effort)
• Active community and regular updates

Use Cases

1. Getting Started Experience

New developers should be able to run:

#Install Ollama
brew install ollama

#Pull a model
ollama pull llama3.2

#Start Lightspeed Stack (no cloud setup needed!)
OLLAMA_MODEL=llama3.2 make run

2. Development & Testing

• Run unit/integration tests without API costs
• Iterate quickly without rate limits
• Test RAG pipelines with local embeddings
• Develop offline

3. CI/CD Pipelines

• Run E2E tests in GitHub Actions without managing secrets
• No quota concerns for parallel test runs
• Reproducible test environment

4. Educational & Demo Purposes

• Workshop attendees don't need cloud accounts
• Demo the system without internet connectivity
• Training environments without budget concerns

5. Privacy-Sensitive Development

• Test with sensitive data locally
• Develop features for air-gapped deployments
• Comply with data sovereignty requirements

Benefits

1. For Developers:

• Zero cost for development
• Faster iteration (no network latency)
• No quota limits during testing
• Lower barrier to contribution

2. For the Project:

• Increased adoption (easier onboarding)
• Better test coverage (developers can test more)
• Broader contributor base (global accessibility)
• Production flexibility (hybrid cloud/local deployments)

3. For Users:

• Try before committing to cloud providers
• Development/staging environments without cloud costs
• Option for fully private deployments

Example Configuration

examples/ollama-run.yaml:

  version: '2'
  image_name: local-development-stack

  providers:
    inference:
      - provider_id: ollama
        provider_type: remote::ollama
        config:
          base_url: http://localhost:11434

  models:
    - model_id: llama3.2
      provider_id: ollama
      model_type: llm
      provider_model_id: llama3.2

    - model_id: nomic-embed-text
      provider_id: ollama
      model_type: embedding
      provider_model_id: nomic-embed-text

Describe alternatives you've considered

1. Continue cloud-only: Creates unnecessary barriers
2. Use sentence-transformers only - Insufficient: Only supports embeddings, not chat
3. Support meta-reference (Llama direct) - More complex setup than Ollama
4. Support multiple local providers - Start with Ollama first, expand later

Additional context

• Similar projects (LangChain, LlamaIndex) prominently feature local model support as a first-class option alongside cloud providers.

• Ollama is already listed in llama-stack's provider list, just needs dependency installation and testing.

Looks like this will be a high-impact, medium-effort improvement that significantly reduces friction for new contributors and developers. The infrastructure is already in place via llama-stack; we primarily need:

• Dependency addition
• Example configurations
• Documentation updates
• Test validation

[tisnik]

@anik120 FYI tracked in Jira https://issues.redhat.com/browse/LCORE-969

[Akrog]

I must be missing something, because I can use ollama just fine:

providers:
  inference:
  - provider_id: ollama
    provider_type: remote::openai
    config:
      base_url: http://localhost:11434/v1/
      api_key: nokey
      allowed_models:
        - "${env.MODEL:=llama3.1:8b}"
      embeddings_metadata: {}
      extra_completion_params:
        drop_params: true
registered_resources:
  models:
  - metadata: {}
    model_id: "${env.MODEL:=llama3.1:8b}"
    provider_id: ollama
    provider_model_id: "${env.MODEL:=llama3.1:8b}"
    model_type: llm

Maybe it's just a matter of documenting it?