ccurme / ghost

Create an AI replica of yourself that is accessible via SMS.

Ghost

Create an AI replica of yourself that is accessible via SMS.

Ghost allows you to chat with (1) a LLM that is fine-tuned on your text messages, or (2) a Langchain agent with a custom prompt that implements retrieval-augmented generation over a set of personal facts. For fine-tuning, Ghost provides functionality to sample and review text conversations from your iPhone, granting oversight of the training data. See fine tuning README for details.

Configuration

Ghost requires two configuration files to tailor your chat responses:

1. Describe your background, writing style, or mannerisms by modifying prompt_prefix.md, or setting the PROMPT_PREFIX_PATH=/path/to/prompt_prefix.md environment variable.
2. Provide structured information such as your SMS-enabled Twilio phone number, chatbot name, and aliases in contacts.json. If using the Langchain agent, this configuration file also houses an array of facts that is indexed and retrieved over when generating responses. You can also point to this configuration file by setting the CONTACTS_PATH=/path/to/contacts.json environment variable.

The contacts.json file specifies other important information such as allowed senders for incoming messages, and information about the senders that helps Ghost tailor its responses. See settings/README.md for more detail.

Usage

Ghost currently uses OpenAI LLMs and embeddings, so you will need to provide an API key. You will also need a Twilio account ID, auth token, and SMS-enabled phone number.

Start server:

TWILIO_ACCOUNT_SID=... TWILIO_AUTH_TOKEN=... OPENAI_API_KEY=... URL=... flask run

or build and run using the Dockerfile.

Ghost has four endpoints:

• /llm_reply: configuring a Twilio SMS webhook to this endpoint will produce replies for inbound messages from known contacts (those in contacts.json).
• /llm_send: this endpoint will send an unsolicited message to a known contact following a given input prompt (requires authentication).
• /llm_memory: this endpoint will fetch conversation histories with all contacts (requires authentication).
• /login: this endpoint can produce a token for use with /llm_send and /llm_memory.

Testing

See Makefile for testing, test coverage and linting.

Unit tests:

make unit_tests

Integration tests:

OPENAI_API_KEY=... make integration_tests