200-line bash script to talk to an LLM prompted with your repo structure + any READMEs it finds (leveraging python and openai)

README-CHAT

#!/bin/bash
# # # # # # # # # # # # # # # # # # # # # # # # # #
# SETTINGS
# # # # # # # # # # # # # #
TARGET_DIR="${1:-./}"  # Default target directory
API_KEY="${2:-$OPENAI_API_KEY}"  # Default API key path
MODEL="${3:-gpt-4}"  # Default model
MAX_DEPTH="${4:-3}"  # Default max depth for tree search
FILE_SIZE_LIMIT_MB="${5:-8}"  # Default file size limit (MB) for the README

# # # # # # # # # # # # # # # # # # # # # # # # # #
# ABOUT
# # # # # # # # # # # # # #
show_help() {
    cat << EOF
Usage: $(basename "$0") [OPTIONS] [TARGET_DIR] [API_KEY] [MODEL] [MAX_DEPTH] [FILE_SIZE_LIMIT_MB]

A single bash script that runs an interactive shell with an LLM prompted on the README files
and file structure of a specified folder (default: ./).

Options:
  -h, --help          Show this help message and exit.

Arguments: (you can set defaults at the top of this script)
  TARGET_DIR          The directory to analyze (default: ./)
  API_KEY             OpenAI API key (default: reads from environment)
  MODEL               LLM model to use (default: gpt-4)
  MAX_DEPTH           Depth for scanning the directory (default: 3)
  FILE_SIZE_LIMIT_MB  Max file size for processing (default: 8MB)

Example:
  $(basename "$0") /path/to/project \$OPENAI_API_KEY gpt-4 3 8
EOF
}

# Check for --help flag
if [[ "$1" == "-h" || "$1" == "--help" ]]; then
    show_help
    exit 0
fi

# # # # # # # # # # # # # # # # # # # # # # # # # #
# PROGRAM
# # # # # # # # # # # # # #
echo "--- README-CHAT ---"
echo "[START]"

# CAPTURE TARGET DIRECTORY STRUCTURE
# -------------------------------------
echo "Saving target directory structure with max depth: $MAX_DEPTH"
TREE_OUTPUT=$(find "$TARGET_DIR" -mindepth 1 -maxdepth "$MAX_DEPTH" -type f -printf "%p\n" | jq -R -s -c 'split("\n")[:-1]') # <-- Capture directory structure with limited depth
README_FILES=$(find "$TARGET_DIR" -mindepth 1 -maxdepth "$MAX_DEPTH" -type f \( -iname "README.md" -o -iname "README.txt" -o -iname "README" \) -printf "%p\n" | jq -R -s -c 'split("\n")[:-1]') # <-- Capture README files

echo "Captured $(echo "$TREE_OUTPUT" | jq '. | length') files in tree structure."
echo "Captured $(echo "$README_FILES" | jq '. | length') README files."

# RUN PYTHON LLM + SHELL
# -------------------------------------
echo "Prompting model..."
$py3 <<EOF
import os, sys, json

import openai

# Load variables from Bash
model = "$MODEL"
api_key = "$API_KEY"
file_structure = json.loads('$TREE_OUTPUT')
readme_files = json.loads('$README_FILES')
max_depth = $MAX_DEPTH
file_size_limit = $FILE_SIZE_LIMIT_MB

# OpenAI client setup
openai.api_key = api_key

# Build optimized prompt for LLM

## Read and concatenate README file contents
def read_readme_files(file_paths):
    readme_content = {}
    for file in file_paths:
        try:
            with open(file, "r", encoding="utf-8") as f:
                readme_content[file] = f.read()
        except Exception as e:
            readme_content[file] = f"[ERROR] Could not read file: {e}"
    return readme_content

## Fetch actual README file contents
readme_contents = read_readme_files(readme_files)

## Prompt
prompt = f"""
You are an AI assistant trained to answer questions about an organization's information repository.

The file organization is as follows:
{json.dumps(file_structure, indent=2)}

The following README files provide key documentation:
{json.dumps(readme_files, indent=2)}

Here are the full contents of the README files:
{json.dumps(readme_contents, indent=2)}

Your task is to help users navigate this information repository and answer their questions accurately.
"""

print(prompt)

# Function to query OpenAI
chat_history = [ # <-- Initialize conversation history
    {"role": "system", "content": prompt}  # System message stays constant
]

def query_llm(user_input):
    client = openai.OpenAI(api_key=api_key)  # Initialize client

    # Append user message to chat history
    chat_history.append({"role": "user", "content": user_input})

    # Send request with history and streaming enabled
    response = client.chat.completions.create(
        model=model,
        messages=chat_history,  # Full conversation history
        stream=True  # Enable streaming
    )

    print("\n[AI]: ", end="", flush=True)  # Show AI label without newline
    ai_response = ""  # Store response for history

    for chunk in response:  # Stream tokens as they arrive
        if chunk.choices[0].delta.content:  # Check if content exists
            token = chunk.choices[0].delta.content
            print(token, end="", flush=True)  # Print token immediately
            ai_response += token  # Store response

    print("\n")  # Newline after response is complete

    # Append AI response to chat history
    chat_history.append({"role": "assistant", "content": ai_response})

# Chat loop
print("\n=== README-CHAT ===")
print("Type your question about the information repository. Type 'exit' to quit.\n")
print("[ --- CHAT --- ]\n")

import sys

# Open a direct connection to the terminal (works even if stdin is detached)
try:
    input_source = open('/dev/tty')
except OSError:
    print("Error: No terminal input available.")
    sys.exit(1)

while True:
    try:
        print(">> ", end="", flush=True)  # Display input prompt
        user_input = input_source.readline().strip()  # Read user input

        if user_input.lower() in ["exit", "quit"]:
            print("Goodbye!")
            break

        query_llm(user_input)  # Stream AI response in real-time

    except EOFError:
        print("\n[INFO] EOF detected. Exiting chat session.")
        break
EOF

echo "[END]"

# # # # # # # # # # # # 
# Example Output:
# ----- 
# === README-CHAT ===
# Type your question about the information repository. Type 'exit' to quit.
#
# [ --- CHAT --- ]
#
# >> what tools should the data scientist use?
#
# [AI]: The Data Scientist is recommended to use the following tools and libraries:
#
# 1. **Python Libraries**
#    - `ibis`: A Python library for building SQL queries.
#    - `sqlalchemy` or `dataset`: These are Python libraries for database operations such as inserting and updating data.
#    - `fastapi`: This is used to serve requests from the web page.
#
# 2. **Database Visualization App**
#    - This could be either Beekeeper Studio or MySQL Workbench. These applications provide a graphical user interface to interact with databases.
#
# 3. `sshfs`: This is used to locally mount the lab drive to one's local machine.|| 
#
# 4. A text editor or Integrated Development Environment (IDE) to write and test the scripts.|| 
#
# 5. An SSH client to connect to the server and execute scripts.|| 
#
# >>