Ghosti Assistant

Ghosti is GhostFilter's focused conversational assistant for scams, phishing, suspicious links, account security, prompt injection, and AI safety. It is an MVP and displays this disclaimer:

Ghosti is still under training, so it can make mistakes. This MVP gives safety guidance, not a guarantee.

Response pipeline

For each relevant message, Ghosti first creates deterministic context from:

• the local scam classifier
• the scam and social-engineering ensemble
• prompt-injection findings
• the GhostGPT agent-firewall decision
• the strongest explainable evidence

When Ollama is reachable, this context is inserted into Ghosti's system prompt and the open model produces the conversational answer. When Ollama is unavailable, Ghosti returns a deterministic answer from the same local evidence.

Unrelated questions are redirected to Ghosti's safety scope. Up to eight recent valid messages are included, and each message is limited to 4,000 characters.

Local Ollama setup

Ollama running on the same machine does not require an API key.

Install Ollama, then run:

ollama serve
ollama pull qwen2.5:3b-instruct

Configure GhostFilter:

GHOSTI_OLLAMA_MODEL=qwen2.5:3b-instruct
OLLAMA_BASE_URL=http://127.0.0.1:11434

Start GhostFilter with npm run dev. The Next.js server calls the local Ollama process; the browser does not connect to Ollama directly.

What judges receive

The deployed Vercel server cannot reach Ollama running on a judge's laptop through 127.0.0.1. Judges can still use Ghosti because the production route returns its deterministic safety fallback.

To test open-model responses, a judge can clone the project and run Ollama locally using the steps above. No Ollama API key is needed for that setup.

Alternatively, a deployment owner can configure a hosted Ollama-compatible endpoint:

OLLAMA_BASE_URL=https://your-protected-model-endpoint.example.com
OLLAMA_API_KEY=your-provider-token
GHOSTI_OLLAMA_MODEL=qwen2.5:3b-instruct

OLLAMA_API_KEY is only for a hosted endpoint that requires bearer authentication. It is not a key that judges must provide to use local Ollama.

Training status

Ghosti has not been presented as a newly fine-tuned model checkpoint. In this MVP, "training" means:

• grounding the model with GhostFilter evidence
• a safety-focused system prompt
• curated examples
• deterministic relevance controls
• regression evaluation and correction feedback

Fine-tuning a consented, redacted dataset and hosting a validated open model are future enhancements.