Skip to contents

Gemini and Ollama

Source: vignettes/providers-gemini-ollama.Rmd
providers-gemini-ollama.Rmd

secure_chat() uses the same provider-neutral path for Gemini, Ollama, and other providers supported by ellmer. All example chunks are unevaluated: building the documentation never installs packages, reads credentials, probes a service, downloads a model, or sends a request.

Install ellmer

Gemini

Store one API key in your user-level ~/.Renviron, not in project files or source control:

GEMINI_API_KEY=replace-with-your-real-key

GOOGLE_API_KEY is also supported. Restart R after editing .Renviron. Do not print, log, or include the key in an audit.

Model availability changes. Inspect the models available to your account, choose one that meets your requirements, and pin its name in production:

This example pins a stable Gemini model. Check the current model list before deploying because availability and retirement dates change:

library(llmshieldr)

gemini_result <- secure_chat(
  prompt = "Explain in one sentence why LLM output should be checked.",
  provider = "gemini",
  model = "gemini-3.8-flash",
  checks = "rules",
  show_stats = TRUE
)

gemini_result[c("action", "output")]

Google controls model availability, quotas, pricing, regions, and data-use terms. Review the current terms and your organization’s data policy before sending private, confidential, or regulated content.

Ollama

Start Ollama and pull a model outside R. For example:

ollama serve
ollama pull gemma3:1b

Pass the model name explicitly so the workflow does not depend on whichever local model happens to be listed first:

library(llmshieldr)

ollama_result <- secure_chat(
  prompt = "Explain in one sentence why LLM output should be checked.",
  provider = "ollama",
  model = "gemma3:1b",
  checks = "rules",
  show_stats = TRUE
)

ollama_result[c("action", "output")]

Use ellmer::models_ollama() interactively to inspect installed models. An Ollama endpoint is local only when its configured base URL is local; review the endpoint, model provenance, and host security before sending sensitive text.

After the basic request works, try one change at a time:

Goal Next step
Review meaning, not only patterns Use checks = "both" with a separately evaluated reviewer.
Guard retrieved documents Add context and a context_policy().
Constrain responses Add an output_contract().
Guard tools Expose only required tools and configure tool_policy().
Inspect usage Set show_tokens = TRUE and show_stats = TRUE.
Evaluate changes Run representative benign, sensitive, and adversarial cases with evaluate_security_cases().

Semantic review may send the same text to a second model or service. Configure reviewer_provider, reviewer_model, and reviewer_provider_args explicitly when that boundary differs from the assistant.

Compatibility Wrappers

shield_gemini() and shield_ollama() are deprecated compatibility wrappers. Use secure_chat(provider = ...) in new code.

Troubleshooting

  • Gemini key not found: confirm the variable is in ~/.Renviron, then restart R.
  • Gemini model or quota error: list models for the active project and check current quota and regional availability.
  • Ollama unavailable: start the service and confirm ollama list contains the exact model passed to secure_chat().
  • Reviewer returns invalid JSON: use a more capable reviewer, test the reviewer contract, and set on_reviewer_error deliberately.