# ml-knowledge-platform

Knowledge verification TUI and KV API client with local LLM support via model-boss.

## Install

```bash
uv add ml-knowledge-platform --extra-index-url https://forge.nasty.sh/api/packages/lilith/pypi/simple/
```

## CLI Usage

```bash
# Launch the TUI
knowledge-platform

# One-shot query
knowledge-platform ask "what is the pricing model?"

# Search the knowledge base
knowledge-platform search "governance"

# Verify a claim
knowledge-platform verify "the platform uses blockchain"

# Check dependencies (KV API + LLM)
knowledge-platform check

# Show active system prompt
knowledge-platform prompt

# List available models
knowledge-platform models
```

## CLI Options

```bash
# Use a specific model
knowledge-platform ask "question" -m claude-haiku

# Override KV API endpoint
knowledge-platform --kv-api http://custom:41233

# Override system prompt
knowledge-platform --system-prompt "You are a custom assistant."

# Skip dependency checks
knowledge-platform --skip-checks

# JSON output
knowledge-platform ask "question" --json-output
knowledge-platform models --json-output
```

## Library Usage

```python
from knowledge_platform import KnowledgeBackend, KVClient, LaunchConfig, ModelConfig

# Use the KV client directly
async with KVClient(base_url="http://localhost:41233") as kv:
    results = await kv.search("pricing model")
    validation = await kv.validate("claim to check")

# Use the full backend (KV + LLM routing)
backend = KnowledgeBackend(kv_api_url="http://localhost:41233")
context = await backend.get_context_for_llm("question")

config = LaunchConfig()
model = config.default_model_object
async for chunk in backend.stream_llm_response(messages, model, context):
    print(chunk, end="")
```

## Environment Variables

| Variable | Description | Default |
|----------|-------------|---------|
| `KNOWLEDGE_SYSTEM_PROMPT` | Override the system prompt | Generic verification prompt |
| `KNOWLEDGE_KV_API_URL` | KV API endpoint | `http://localhost:41233` |
| `ANTHROPIC_API_KEY` | For Claude models | — |

## Local LLM Infrastructure

Default models use local inference via `model-boss` + `llama-http`:

- **Port 10010**: Ministral 3B (fast formatting) via `llama-http-3b.service`
- **Port 10020**: Ministral 14B (reasoning) via `llama-http-14b.service`

The platform auto-starts systemd user services when needed and manages GPU VRAM leases through model-boss.

## Custom System Prompts

Inject domain-specific knowledge at runtime:

```bash
# Via environment variable
export KNOWLEDGE_SYSTEM_PROMPT="You are Crystal, the Lilith platform expert..."
knowledge-platform

# Via CLI flag
knowledge-platform --system-prompt "You are a medical knowledge assistant..."

# Via config file (~/.config/knowledge-platform/config.toml)
echo 'system_prompt = "Your custom prompt here"' > ~/.config/knowledge-platform/config.toml
```

Priority: CLI flag > env var > config file > built-in default.