LLM Usage Patterns

Documentation status: reference — see Maturity and evidence.

Use LLMs as governed inference components around conceptual memory rather than as the sole owner of application truth.

Useful patterns include retrieval/context assembly, explanation, model-authoring assistance, hypothesis generation and agent/tool orchestration. Validate generated actions or model changes against Runtime rules and governance boundaries.

For local inference, choose deliberately between the native logiCells GGUF/Transformer path and the provider-neutral LLM service path. The llama.cpp provider is appropriate when the application wants a resident local llama.cpp model behind the generic service contract.

Keep these boundaries explicit:

• conceptual retrieval selects and identifies authoritative context;
• LLM inference proposes or generates text/structured output;
• symbolic/runtime validation decides what can become an accepted fact, action or model change;
• provenance records which model/provider/configuration produced the proposal.

See Run a local model with the llama.cpp provider for the local provider lifecycle.