ai-assistant/assistant.py

#!/bin/python3
# Chat with an intelligent assistant in your terminal
from ollama import Client
import re
import pyperclip
import sys
import argparse
import pygments
from pygments.lexers import get_lexer_by_name
from pygments.formatters import TerminalFormatter


server = 'localhost:11434'
model = 'llama3.1:8b-instruct-q8_0'
temp = 0.2

pattern = r'```[a-z]*\n[\s\S]*?\n```'
line_pattern = r'`[a-z]*[\s\S]*?`'

def parse_commands(text):
    # See if user wrote any commands here
    tokens = text.split(' ')
    match tokens[0]:
        case '/save':
            print('saving')
        case '/clear':
            print('clearing context')
        case '/exit':
            print('exiting')

def highlight_code(language_name, code):
    # Check if the language is specified in the first line
    lexer_name = language_name
    if lexer_name == None:
        lines = code.split('\n')
        for line in lines:
            if line.strip().startswith('```'):
                lexer_name = line.strip().split('```')[1].strip()
                break
            elif line.strip().startswith('lang:'):
                lexer_name = line.strip().split(':')[1].strip()
                break

    if lexer_name:
        try:
            # Try to get the lexer by name
            lexer = get_lexer_by_name(lexer_name)
        except ValueError:
            # If the lexer is not found, guess it
            lexer = guess_lexer(code.split('\n')[1:-1])
            if not lexer:
                # If no lexer is guessed, default to bash
                lexer = get_lexer_by_name('bash')
    else:
        # If no language is specified, guess the lexer
        print("LEXER NAME " + lexer_name)
        lexer = guess_lexer(code.split('\n')[1:-1])
        if not lexer:
            # If no lexer is guessed, default to bash
            lexer = get_lexer_by_name('bash')

    formatter = TerminalFormatter()

    just_code = code.split('\n')[0]
    newlines = '\n'.join(code.split('\n')[1:])
    # if code is a code block, strip surrounding block markers
    lines = code.split('\n')
    if (len(lines) > 2) and ('```' in lines[0]) and ('```' in lines[-1]):
        just_code = '\n'.join(code.split('\n')[1:-1])

    highlighted_code = pygments.highlight(just_code, lexer, formatter)
    return highlighted_code + newlines

def extract_code_block(markdown_text):
    # Use the regular expression pattern to find all matches in the markdown text
    matches = re.finditer(pattern, markdown_text)

    # Iterate over the matches and extract the code blocks
    code_blocks = []
    for match in matches:
        code_block = match.group(0)

        highlighted_code = highlight_code(None, code_block)

        # Add the highlighted code block to the list of code blocks
        code_blocks.append(highlighted_code)


    if len(code_blocks) == 0:
        line_matches = re.finditer(line_pattern, markdown_text)
        for match in line_matches:
            code_block = match.group(0)
            code_blocks.append(code_block[1:-1])
    return code_blocks

def copy_string_to_clipboard(string):
    try:
        pyperclip.copy(string)
    except:
        return


code_history = []

history = [
        {"role": "system", "content": "You are a helpful, smart, kind, and efficient AI assistant. You always fulfill the user's requests accurately and concisely."},
        ]

conversation = ""

def chat(message, stream=True):
    history.append({"role": "user", "content": message})
    completion = client.chat(
            model=model,
            options={"temperature":temp},
            messages=history,
            stream=stream
            )
    result = ''
    language = ''
    large_chunk = []
    for chunk in completion:
        if stream:
            text = chunk['message']['content']
            large_chunk.append(text)
            large_text = ''.join(large_chunk)
            
            # update language if entering or leaving code block
            if ('\n' in large_text) and ('```' in large_text):
                language = large_text.split('```')[1].split('\n')[0]
                if language == '':
                    language = None
                print(large_text, end='', flush=True)
                large_chunk = []
                large_text = ''

            # Only print full lines
            if '\n' in large_text:
                output = large_text
                if language:
                    output = highlight_code(language, output)
                print(output, end='', flush=True)
                large_chunk = []
            result += text
    if not stream:
        result = completion['message']['content']
    if stream:
        print(large_text, flush=True)
    history.append({"role": 'assistant', 'content': result})
    return result

def chat2(args, user_input, stream=True):
    conversation += 'user: ' + user_input + '\n'
    if args.reflect:
        result = reflection_mode(user_input, stream)
    else:
        result = chat(user_input, stream)
    conversation += 'assistant: ' + result + '\n'
    return result

def highlightify_text(full_text):
    lines = full_text.split('\n')
    result = ''
    language = None
    for line in lines:
        text = line + '\n'
        
        # update language if entering or leaving code block
        if '```' in text:
            language = text.split('```')[1].split('\n')[0]
            if language == '':
                language = None
            result += text
            text = ''

        # Only print full lines
        if '\n' in text:
            output = text
            if language:
                output = highlight_code(language, output)
            result += output
    return result

def parse_args():
    # Create the parser
    parser = argparse.ArgumentParser(description='Copy and open a source file in TextEdit')
    # Add the --follow-up (-f) argument
    parser.add_argument('--follow-up', '-f', nargs='?', const=True, default=False, help='Ask a follow up question when piping in context')
    # Add the --copy (-c) argument
    parser.add_argument('--copy', '-c', action='store_true', help='copy a codeblock if it appears')
    # Add the --shell (-s) argument
    parser.add_argument('--shell', '-s', nargs='?', const=True, default=False, help='output a shell command that does as described')
    # Add the --model (-m) argument
    parser.add_argument('--model', '-m', nargs='?', const=True, default=False, help='Specify model')
    # Add the --temp (-t) argument
    parser.add_argument('--temp', '-t', nargs='?', const=True, default=False, help='Specify temperature')
    # Add the --host argument
    parser.add_argument('--host', nargs='?', const=True, default=False, help='Specify host of ollama server')
    # Add the --reflect argument
    parser.add_argument('--reflect', action='store_true', help='Use reflection prompting style to improve output. May be slower and not work with all models.')
    # Parse the arguments
    return parser.parse_args()

def reflection_mode(query, should_print=False):
    reflection_prompt = """
You are a helpful ai assistant that answers every question thoroughly and accurately. You always begin your response with a <planning></planning> section where you lay out your plan for answering the question. It is important that you don't make any assumptions while planning. Then you <reflect></reflect> on your plan to make sure it correctly answers the user's question. Then, if you are confident your plan in correct, you give your <draft answer>, followed by <final reflection> to make sure the answer correctly addresses the user's question. Finally, give a <final answer> with your answer to the user. If there are any ambiguous or unknown requirements, ask the user for more information as your final answer. You must always have a <final answer> no matter what, even if you are asking for clarifying questions. If you do not have the <final answer> tags, the user will not see your response. Additionally, the user can not see your planning or reflecting, they can only see what goes in the <final answer></final answer> tags, so make sure you provide any information you want to tell the user in there.


"""
    result = chat(reflection_prompt + query, stream=False)
    highlighted_result = highlightify_text(result)

    # print('==DEBUG==')
    # print(highlighted_result)
    # print('==DEBUG==')

    
    final_answer = highlighted_result.split('<final answer>')
    while len(final_answer) < 2:
        final_answer = chat('Please put your final answer in <final answer></final answer> tags.', stream=False)
        final_answer = highlighted_result.split('<final answer>')
    final_answer = final_answer[1].split('</final answer>')[0]

    if should_print:
        print(final_answer)
    return final_answer

def set_host(host):
    global server
    server = host

def arg_follow_up(args):
    sys.stdin = open('/dev/tty')
    if args.follow_up != True:
        second_input = args.follow_up
    else:
        second_input = input('> ')
    return second_input

def arg_shell(args):
    query = 'Form a shell command based on the following description. Only output a working shell command .\nDescription: '
    if args.shell != True:
        query += args.shell
    else:
        query += input('> ')
    result = chat2(args, query, False)
    result = blocks[0] if len(blocks := extract_code_block(result)) else result
    print(result)
    copy_string_to_clipboard(result)

def handle_piped_input(args):
    all_input = sys.stdin.read()
    query = 'Use the following context to answer the question. There will be no follow up questions from the user so make sure your answer is complete:\nSTART CONTEXT\n' + all_input + '\nEND CONTEXT\nAfter you answer the question, reflect on your answer and determine if it answers the question correctly.'
    if args.copy:
        query += 'Answer the question using a codeblock for any code or shell scripts\n'
    if args.follow_up:
        query += arg_follow_up(args)
    query += '\n'

    result = chat2(args, query)
    blocks = extract_code_block(result)
    if args.copy and len(blocks):
        copy_string_to_clipboard(blocks[0])

def handle_non_piped_input(args):
    if args.shell:
        arg_shell(args)
        exit()
    if args.follow_up:
        user_input = arg_follow_up(args)
        result = chat2(args, user_input)
        exit()
    while True:
        try:
            user_input = input('> ')
        except (EOFError, KeyboardInterrupt):
            print()
            exit()
        else:
            result = chat2(args, user_input)

client = None

def main():
    args = parse_args()
    if args.host:
        set_host(args.host)
    # Point to the local server
    global client
    client = Client(host=server)
    if args.model:
        global model 
        model = args.model
    if args.temp:
        global temp
        temp = float(args.temp)
    if not sys.stdin.isatty():
        handle_piped_input(args)
    else:
        handle_non_piped_input(args)

if __name__ == '__main__':
    main()