Invoke LLMs from the command line

ai.sh

#!/usr/bin/env bash
#
# ai.sh - invoke llm from command line
#
# Usage:
# * Run interactively
#   ai.sh
#
# * Invoke a single inference from command line
#   ai.sh ffmpeg convert all mp4 in this folder to mp3
#
# * Inject text content into the context
#   cat document.txt | ai summarize this document
#
# * Inject text content into interactive session
#   cat document.txt | ai
#

aish_openai_endpoint='https://openrouter.ai/api/v1'
aish_openai_key='sk-or-v1-xxxxxxxxxxxxxxxx'
aish_openai_model='anthropic/claude-3.5-sonnet'
system='You are an assistant invoked through a command line. Generate a code
snippet to fulfill the task. By default: only output the command, no explanations,
unless otherwise asked by user. If asked to explain, format responses such that
it looks good in the CLI, e.g. hard wrap to 80 characters, and be concise.'

require() {
	cmd="$1"
	if ! command -v "$cmd" >/dev/null; then
		echo "$cmd not installed"
		exit 1
	fi
}

require curl
require jq

# Attach stdin if it exists
prompt="$@"
interactive=false
extra_content=''
if [[ ! -t 0 ]]; then
	stdin=`cat`
	extra_content="[stdin after this line]\n$stdin"
	nchars=$(wc -c <<< "$stdin")
	echo "[*] Loaded $nchars bytes from stdin into context"
fi

messages=$(jq -n \
	--arg system "$system" \
	--arg extra "$extra_content" \
	'[
		{ role: "system", content: $system },
		(if $extra != "" then
			{ role: "user", content: $extra }
		else
			empty
		end)
	]'
)

add_message() {
	role="$1"
	content="$2"

	messages=$(jq \
		--arg role "$role" \
		--arg content "$content" \
		'. + [ { role: $role, content: $content } ]' \
		<<< "$messages"
	)
}

inference() {
	# Prompt user if there is no prompt from the arguments
	if [[ -z "$prompt" ]]; then
		interactive=true
		IFS= read -r -p "> " prompt < /dev/tty || {
			echo '^D'
			exit 0
		}
	fi

	# Update context
	add_message user "$prompt"

	# Construct request object
	req_json=$(jq -n \
		--arg model "$aish_openai_model" \
		--argjson messages "$messages" \
		'{
			model: $model,
			messages: $messages,
			stream: true
		}'
	)

	# Send streaming request
	curl -fsSLN "$aish_openai_endpoint/chat/completions" \
		-XPOST \
		-H "Content-Type: application/json" \
		-H "Authorization: Bearer $aish_openai_key" \
		-d "$req_json" \
		| grep --line-buffered -E '^data: ' \
		| grep --line-buffered -v '[[DONE]]' \
		| stdbuf -oL cut -c 7- \
		| jq --unbuffered -j '.choices[0].delta.content' \
		| tee >/dev/stderr | read -r response
	echo ''

	# Update context
	add_message assistant "$response"

	# Prepare for next loop if any
	prompt=''
}

inference
while [ "$interactive" = true ]; do
	inference
done