An AI command line chat bot

An AI chat bot you can run from the command line.

By default, it provides a proxy onto the underlying Google Gemini APIs, which is available on a free tier. When you start the chat bot you provide a context, either an opening question, file inputs or a prompt template. From that point you can ask further questions from that context.

ask-voice-session-2.mov

If you set up an voice to text transcribe service then you can control the command line with voice. See below for details.

Installation

You can either run direct from source

pip3 install -r requirements.txt
python ask/ask.py hello

Or install globally

pip install .
ask hello

Or install globally in editable mode, for example if you want to make modifications to the code

pip install -e .
ask hello

Getting started

You'll need a Gemini API Key to get started. You can get one from the Google AI Studio. Set the key as an environment variable.

export GEMINI_API_KEY=my-super-secret-key

Then start a chat

ask

You can pass the contents of a file in as an input

ask file.txt

Or ask a question to work on

ask what time is love

Hit return to ask the default question, e.g. proof read or answer the question. Or elaborate and ask more.

Voice to text

Provide inputs from voice by setting up a text service that outputs transcribed text to the file /tmp/transcribe.txt.

For example build whisper.cpp and run in another shell.

git clone https://github.com/ggerganov/whisper.cpp.git
cd whisper.cpp
bash ./models/download-ggml-model.sh small.en
make
./stream -m models/ggml-large-v3-q5_0.bin --length 1000 --keep 0 -f /tmp/transcribe.txt

Example

Start off a chat on a given theme and then quick fire requests

❯ ask show me terraform for the following

(-_-) s3

resource "aws_s3_bucket" "example" {
  bucket = "my-bucket-name"

  acl = "private"

  tags = {
    Name = "My bucket"
  }
}

(-_-) vpc

resource "aws_vpc" "main" {
  cidr_block = "10.0.0.0/16"

  enable_dns_hostnames = true
  enable_dns_support = true
}

resource "aws_subnet" "public" {
  vpc_id = aws_vpc.main.id
  cidr_block = "10.0.1.0/24"
  map_public_ip_on_launch = true
}

(-_-) rds

resource "aws_db_instance" "default" {
  allocated_storage = 10
  engine             = "mysql"
  engine_version     = "5.7.28"
  instance_class     = "db.t2.micro"
  name               = "mydb"
  username           = "admin"
  password           = "password"
  db_subnet_group_name = "default"
}

resource "aws_db_subnet_group" "default" {
  name       = "default"
  subnet_ids = [aws_subnet.public.id]
}

Prompt templates

You can build up a library of prompt templates for your use cases.

For example you can start with the provided lexical.txt prompt as an example.

ask --template lexical.txt dog poodle fish

This prompt template has {} placeholders, e.g.

You are skilled lexical scanner. Can you tell me whether the phrase {} relates to phrase 1 or phrase 2 below
Phrase 1: {}
Phrase 2: {}

You should get content like:

The phrase "dog" relates to **phrase 1: poodle**.
"Poodle" is a type of dog, making them both related.  "Fish" is a completely different animal category.

You can create a prompt directory for your collection of personal of prompts and then set this directory with the env variable name ASK_PROMPT_DIRECTORY. Each prompt should be defined in a .txt files in that directory. Then refer to the template with the base name. For example the following would use my-prompt in the ASK_PROMPT_DIRECTORY directory.

ask --template my-prompt one two three

One shot question and answer

Ask a one shot question and expect a single answer

ask --one explain relativity to me in 100 words

You can enable this mode in the shell using the ASK_PIPELINE environment variable.

export ASK_PIPELINE=one-shot
ask can you describe Falmouth
ask pi to 100 decimal places