llm_chat ยท diff
git:20260902.af2a47c to git:20260906.f9cef19
20 added, 25 removed. Audit A to A.
---
name: llm_chat
- description: Ask the configured LLM_MODEL a question using user-provided context, questions, and answer instructions. Use when a task needs a direct LLM response over supplied context, typically context gathered from semantic search, with explicit rules or restrictions for the answer.
+ description: Test the configured text-generation model with a supplied prompt. Use for direct provider checks without dataset retrieval.
---
- # LLM Chat Skill
+ # LLM chat
- A small command for testing the configured text-generation model. Runtime code
- uses the infrastructure `generate_markdown` and `generate_json` interfaces.
+ Render a direct response from the configured text-generation service.
- ## Setup
+ ## Operations and effects
- 1. All runtime dependencies (`typer`, `python-dotenv`, `rich`, `litellm`) are installed in the `sictic-env` Conda environment by `{{REPO_ROOT}}/install.sh`.
- 2. The `.env` file at `{{REPO_ROOT}}/.env` must define:
- - `LLM_MODEL`: The default text-generation model to use (e.g., `gemini/gemini-1.5-pro`, `ollama/llama3`).
- - `LLM_BASE_URL`: Optional endpoint base URL. Use `http://localhost:11434` for local Ollama.
- - `LLM_API_KEY`: Optional endpoint API key. Leave blank for local Ollama.
- - `GEMINI_API_KEY`: Your Google Gemini API key.
- - `OLLAMA_HOST`: The host URL for Ollama (if using Ollama).
- - `OLLAMA_CONTEXT_LENGTH`: The baseline token context for Ollama.
+ This is a CLI wrapper around `generate_markdown`, not a skill-level generation
+ API. Other workflows call the shared generation service directly. It has no
+ harness command or bulk-refresh registration.
- ## Usage
+ The command calls the configured model and renders Markdown through Rich.
+ It performs no retrieval or insight persistence. Model/configuration failures
+ use the shared CLI error handling; scheduling and recovery belong to the
+ generation service.
- `llm_chat` is not exposed as a harness slash command. User-facing harness commands
- call it internally through the shared skill APIs. Run it directly only when testing
- provider configuration or debugging model behavior.
+ Configure the model, endpoint and applicable credentials through the
+ [installation guide](../../docs/installation-and-operations.md). A Gemini key is
+ not required when using another provider.
- ```bash
- # Basic usage with default model
- conda run -n sictic-env python -m skills.llm_chat "What is startup due diligence?"
+ ## Usage
+ ```bash
+ conda run -n sictic-env python -m skills.llm_chat "Explain startup due diligence."
```
- ## Features
+ ## References
- - **Provider-neutral generation:** Uses the shared AI text-generation infrastructure
- - **Ollama context management:** Sizes context between `OLLAMA_CONTEXT_LENGTH` and `OLLAMA_CONTEXT_LENGTH_MAX`
- - **Rich Formatting:** Console outputs and markdown are beautifully rendered using the `Rich` library.
- - **Shared recovery:** Uses the standard provider-error handling and three-attempt limit
+ - [CLI implementation](__main__.py)
+ - [Shared generation](../standards_and_architecture/SKILL.md#model-calls)