Milla is an IRC bot that:
- sends things over to an LLM when you ask it questions and prints the answer with optional syntax-highlighting.Currently supported providers: Ollama, Openai, Gemini, Openrouter
- Milla can run more than one instance of itself
- Each instance can connect to a different ircd, and will get the full set of configs, e.g. different proxies, different postgres instance, ...
- You can define custom commands in the form of SQL queries to the database with the SQL query result being passed to the bot along with the given prompt and an optional limit so you don't go bankrupt(unless you are running ollama locally like the smart cookie that you are).
- lua plugin system to extend the bot's functionality.
milla accepts one cli arg which tells it where to look for the config file:
Usage of ./milla:
-config string
path to the config file (default "./config.toml")
The bot will respond to chat prompts if they begin with botnick:.
The bot will see a chat prompt as a command if the message begins with botnick: /.
An example is provided under config-example.toml. Please note that all the config options are specific to one instance which is defined by ircd.nameofyourinstance.
The address for the IRC server to connect to.
The nick the bot should use.
Whether to use SASL for authentication.
The SASL username.
The SASL password for SASL plain authentication. Can also be passed as and environment variable.
The address for the Ollama chat endpoint.
The name of the model to use.
The style to use for syntax highlighting done by chroma. This is basically what's called a "theme".
The formatter to use. This tells chroma how to generate the color in the output. The supported options are:
noopfor no syntax highlightingterminalfor 8-color terminalsterminal8for 8-color terminalsterminal16for 16-color terminalsterminal256for 256-color terminalsterminal16mfor truecolor terminalshtmlfor HTML output
NOTE: please note that the terminal formatters will increase the size of the IRC event. Depending on the IRC server, this may or may not be a problem.
Which LLM provider to use. The supported options are:
- ollama
- chatgpt
- gemini
- openrouter
The apikey to use for the LLM provider. Can also be passed as and environment variable.
The system message to use for ollama.
The path to the client certificate to use for client cert authentication.
The password to use for the IRC server the bot is trying to connect to if the server has a password. Can also be passed as and environment variable.
Which address to bind to for the IRC server.
The temperature to config the model with.
The timeout for requests made to the LLM provider.
How much to wait before reconnecting to the IRC server.
Which port to connect to for the IRC server.
How many conversations to keep in memory for a model.
Ping delay for the IRC server.
Ping timeout for the IRC server.
Skip verifying the IRC server's TLS certificate. This only makes sense if you are trying to connect to an IRC server with a self-signed certificate.
Whether to use TLS to connect to the IRC server. This option is provided to support usage on overlay networks such as Tor, i2p and yggdrassil.
Disables the "fallback" to a non-TLS connection if the strict transport policy expires and the first attempt to reconnect back to the TLS version fails.
Disable girc's built-in flood protection.
Whether to enable debug logging. The logs are written to stdout.
Whether to write raw messages to stdout.
List of admins for the bot. Only admins can use commands.
admins = ["admin1", "admin2"]
List of channels for the bot to join when it connects to the server.
ircChannels = [["#channel1","channel1password"], ["#channel2",""], ["#channel3"]]
In the provided example, milla will attempt to join #channel1 with the provided password while for the other two channels, it will try to join normally.
This behaviour is consistant across all places where a channel name is the input.
Please note that the bot does not have to join a channel to be usable. One can simply query the bot directly as well.
Name of the database user.
Password for the database user.
Address of the database.
Name of the database.
List of channels that the bot will scrape into a database table. You can later on use these databases for the custom commands.
ircChannels = [["#channel1","channel1password"], ["#channel2",""], ["#channel3"]]
Determines which proxy to use to connect to the IRC network:
ircProxy = "socks5://127.0.0.1:9050"
Determines which proxy to use to connect to the LLM endpoint:
llmProxy = "socks5://127.0.0.1:9050"
Determines which proxy to use for other things:
llmProxy = "socks5://127.0.0.1:9050"
Name of the milla instance, must be unique across all instances.
Milla will only answer if the nick is in the admin list.
webirc gateway to use.
webirc hostname to use.
webirc password to use.
webirc address to use.
the context to use for the normal conversations with the bot. Yes, this is how you tell your milla instance to act like a pirate.
context = ["you are a pirate. use the language and words a pirate would unless you are asked to do otherwise explicitly", "your name is caption blackbeard"]context = ["please respond in french even if i use another language unless you are specifically asked to use any language other than french", "your name is terra"]The file that contains the rss feeeds.
The channel to send the rss feeds to.
A list of plugins to load:plugins = ["./plugins/rss.lua", "./plugins/test.lua"]
Custom commands let you define a command that does a SQL query to the database and performs the given task. Here's an example:
[ircd.devinet_terra.customCommands.digest]
sql = "select log from liberanet_milla_us_market_news order by log desc;"
limit = 300
context = ["you are a sentiment-analysis bot"]
prompt= "i have provided to you news headlines in the form of previous conversations between you and me using the user role. please provide the digest of the news for me."
[ircd.devinet_terra.customCommands.summarize]
sql= "select log from liberanet_milla_us_market_news order by log desc;"
limit= 300
context = ["you are a sentiment-analysis bot"]
prompt= "i have provided to you news headlines in the form of previous conversations between you and me using the user role. please summarize the provided news for me. provide some details."
[ircd.devinet_terra.customCommands.canada]
sql= "select log from liberanet_milla_us_market_news order by log desc;"
limit= 300
context = ["you are a canadian news anchor", "you only care about news that is relevant to canada"]
prompt= "i have provided to you news headlines in the form of previous conversations between you and me using the user role. please summarize the provided news for me. provide some details."In the above example digest and summarize will be the names of the commands: milla: /cmd summarize.
Currently you should only ask for the log column in the query. Asking for the other column will result in the query not succeeding.
The limit parameter limits the number of SQL queries that are used to generate the response. Whether you hit the token limit of the provider you use and the cost is something you should be aware of.
A limit value of 0 disables the limit on the amount of rows that are passed to milla.
NOTE: since each milla instance can have its own database, all instances might not necessarily have access to all the data milla is gathering. If you use the same database for all the instances, all instances will have access to all the gathered data.
Watchlists allow you to specify a list of channels to watch. The watched values are given in a list of files, each line of the file specifying a value to watch for. Finally a value is given for the alertchannel where the bot will mirror the message that triggered a match.
[ircd.devinet_terra.watchlist.security]
watchList = [["#securityfeeds"]]
watchFiles = ["/watchfiles/voidbox.list"]
alertChannel = ["#milla_alerts"]
eventTypes = ["PRIVMSG"]
fgColor = 0
bgColor = 28For the watchList option, please remember to put the channels also in your scrape scrapeChannels.
The rss file is self-explanatory. Here's an example:
{
"feeds": [
{
"name": "one",
"url": "http://feeds.feedburner.com/crunchyroll/rss",
"proxy": "socks5://172.17.0.1:9007",
"userAgent": "Mozilla/5.0 (X11; U; Linux i686; pl; rv:1.8.1.1) Gecko/20061204 Firefox/2.0.0.1 (Ubuntu-edgy)",
"timeout": 10
},
{
"name": "two",
"url": "http://feeds.feedburner.com/crunchyroll/rss/anime",
"proxy": "socks5://172.17.0.1:9007",
"userAgent": "Mozilla/5.0 (X11; U; Linux i686; pl; rv:1.8.1.1) Gecko/20061204 Firefox/2.0.0.1 (Ubuntu-edgy)",
"timeout": 10
}
],
"period": 3600
}[ircd.devinet]
ircServer = "irc.myawesomeircnet.com"
ircPort = 6697
ircNick = "milla"
enableSasl = true
ircSaslUser = "milla"
ircSaslPass = "xxxxx"
ircChannels = [["##chan1"], ["##chan2"]]
temp = 0.2
requestTimeout = 10
millaReconnectDelay = 60
model = "gpt-3.5-turbo"
chromaStyle = "rose-pine-moon"
chromaFormatter = "terminal256"
provider = "chatgpt"
apikey = "xxxx"
memoryLimit = 20
admins = ["noone_has_this_nick"]
debug = true
out = true
databaseAddress = "postgres:5432"
databasePassword = "changeme"
databaseUser = "devi"
databaseName = "milla"
scrapeChannels = [["#soulhack"], ["#warroom"], ["#securityfeeds"]]
ircProxy = "socks5://127.0.0.1:9050"
llmProxy = "http://127.0.0.1:8180"
skipTLSVerify = false
useTLS = true
adminOnly = false
plugins = ["/plugins/ip.lua", "/plugins/urban.lua"]
context = ["please respond in french even if i use another language unless you are specifically asked to use any language other than french"]
[ircd.devinet.watchlist.security]
watchList = [["#securityfeeds"]]
watchFiles = ["/watchfiles/voidbox.list"]
alertChannel = ["#milla_alerts"]
eventTypes = ["PRIVMSG"]
fgColor = 0
bgColor = 28
[ircd.devinet.rss.manga]
rssFile = "/rssfeeds/manga.json"
channel = ["#manga"]
[ircd.devinet.rss.anime]
rssFile = "/rssfeeds/anime.json"
channel = ["#anime"]
[ircd.liberanet]
ircServer = "irc.libera.chat"
ircNick = "milla"
model = "gpt-3.5-turbo"
ircPort = 6697
chromaStyle = "rose-pine-moon"
chromaFormatter = "terminal16m"
provider = "gemini"
apikey = "xxxx"
temp = 0.5
requestTimeout = 10
millaReconnectDelay = 60
keepAlive = 20
memoryLimit = 20
pingDelay = 20
pingTimeout = 600
skipTLSVerify = false
useTLS = true
disableSTSFallback = true
allowFlood = false
admins = ["noone_has_this_nick"]
ircChannels = [["##milla1"], ["##milla2"]]
debug = true
out = true
ircProxy = "socks5://127.0.0.1:9051"
llmProxy = "http://127.0.0.1:8181"
adminOnly = true
[ircd.liberanet.customCommands.digest]
sql = "select log from liberanet_milla_us_market_news order by log desc;"
limit = 300
context = ["you are a sentiment-analysis bot"]
prompt= "i have provided to you news headlines in the form of previous conversations between you and me using the user role. please provide the digest of the news for me."
[ircd.liberanet.customCommands.summarize]
sql= "select log from liberanet_milla_us_market_news order by log desc;"
limit= 300
context = ["you are a sentiment-analysis bot"]
prompt= "i have provided to you news headlines in the form of previous conversations between you and me using the user role. please summarize the provided news for me. provide some details."Prints the help message.
Get the value of a config option. Use the same name as the config file but capitalized: /get chromaFormatter
Get the value of all config options.
Set a config option on the fly. Use the same name as the config file but capitalized: /set chromaFormatter noop
Returns memory stats for milla.
Joins a channel: /join #channel [optional_password]
Leaves a channel: /leave #channel
Load a plugin: /load /plugins/rss.lua
Unload a plugin: /unload /plugins/rss.lua
Pings the user after the given amount in seconds: /remind 1200
Rolls a number between 1 and 6 if no arguments are given. With one argument it rolls a number between 1 and the given number. With two arguments it rolls a number between the two numbers: /rool 10000 66666
IANA whois endpoint query: milla: /whois xyz
Images are automatically pushed to dockerhub. So you can get it from there.
An example docker compose file is provided in the repo under docker-compose.yaml.
milla can be used with gvisor's docker runtime, runsc.
services:
terra:
image: milla_distroless_vendored
build:
context: .
dockerfile: ./Dockerfile_distroless_vendored
deploy:
resources:
limits:
memory: 128M
logging:
driver: "json-file"
options:
max-size: "100m"
networks:
- terranet
user: 1000:1000
restart: unless-stopped
entrypoint: ["/usr/bin/milla"]
command: ["--config", "/config.toml"]
volumes:
- ./config-gpt.toml:/config.toml
- /etc/localtime:/etc/localtime:ro
cap_drop:
- ALL
postgres:
image: postgres:16-alpine3.19
deploy:
resources:
limits:
memory: 4096M
logging:
driver: "json-file"
options:
max-size: "200m"
restart: unless-stopped
ports:
- "127.0.0.1:5455:5432/tcp"
volumes:
- terra_postgres_vault:/var/lib/postgresql/data
- ./scripts/:/docker-entrypoint-initdb.d/:ro
environment:
- POSTGRES_PASSWORD_FILE=/run/secrets/pg_pass_secret
- POSTGRES_USER_FILE=/run/secrets/pg_user_secret
- POSTGRES_INITDB_ARGS_FILE=/run/secrets/pg_initdb_args_secret
- POSTGRES_DB_FILE=/run/secrets/pg_db_secret
networks:
- terranet
- dbnet
secrets:
- pg_pass_secret
- pg_user_secret
- pg_initdb_args_secret
- pg_db_secret
runtime: runsc
pgadmin:
image: dpage/pgadmin4:8.6
deploy:
resources:
limits:
memory: 1024M
logging:
driver: "json-file"
options:
max-size: "100m"
environment:
- PGADMIN_LISTEN_PORT=${PGADMIN_LISTEN_PORT:-5050}
- PGADMIN_DEFAULT_EMAIL=${PGADMIN_DEFAULT_EMAIL:
- PGADMIN_DEFAULT_PASSWORD_FILE=/run/secrets/pgadmin_pass
- PGADMIN_DISABLE_POSTFIX=${PGADMIN_DISABLE_POSTFIX:-YES}
ports:
- "127.0.0.1:5050:5050/tcp"
restart: unless-stopped
volumes:
- terra_pgadmin_vault:/var/lib/pgadmin
networks:
- dbnet
secrets:
- pgadmin_pass
networks:
terranet:
driver: bridge
dbnet:
volumes:
terra_postgres_vault:
terra_pgadmin_vault:
secrets:
pg_pass_secret:
file: ./pg/pg_pass_secret
pg_user_secret:
file: ./pg/pg_user_secret
pg_initdb_args_secret:
file: ./pg/pg_initdb_args_secret
pg_db_secret:
file: ./pg/pg_db_secret
pgadmin_pass:
file: ./pgadmin/pgadmin_passThe env vars UID and GID need to be defined or they can replaces by your host user's uid and gid.
As a convenience, there is a distroless dockerfile, Dockerfile_distroless also provided.
A vendored build of milla is available by first running go mod vendor and then using the provided dockerfile, Dockerfile_distroless_vendored.
For a regular build:
go mod download
go buildFor a vendored build:
go mod vendor
go buildmilla can be extended with plugins. The plugins are written in lua and are loaded at runtime. The plugins are loaded after an IRC connection has been made.
milla uses gopher-lua which implements a lua 5.1 VM in Go.
This means that lua libraries that are implemented in C will not be available to gopher-lua, only pure lua libraries will be available..
There are a few libraries written in go specifically for gopher-lua that are available to milla. Below there is a list of the current ones.
An example plugin is provided under plugins/rss.lua.
period: 3600
channel: "#rssfeed"
rssfeeds:
- name: "one"
url: "https://www.youtube.com/feeds/videos.xml?channel_id=UCaiL2GDNpLYH6Wokkk1VNcg"
- name: "two"
url: "https://www.youtube.com/feeds/videos.xml?channel_id=UCd26IHBHcbtxD7pUdnIgiCw"
- name: "three"
url: "https://www.youtube.com/feeds/videos.xml?channel_id=UCS4FAVeYW_IaZqAbqhlvxlA"local milla = require("milla")
local yaml = require("yaml")
local http = require("http")
local xmlpath = require("xmlpath")
local function read_file(file)
local f = assert(io.open(file, "rb"))
local content = f:read("*all")
f:close()
return content
end
local function sleep(n) os.execute("sleep " .. tonumber(n)) end
local function get_config()
local yaml_config = read_file("./plugins/rss.yaml")
local config = yaml.parse(yaml_config)
return config
end
local function get_rss_feed(config)
local titles = {}
local author_names = {}
local uris = {}
local rss_feed_list = {}
for _, v in pairs(config.rssfeeds) do
local response, err = http.request("GET", v.url)
if err ~= nil then
milla.send_message(err, "")
goto continue
end
local node, err = xmlpath.loadxml(response.body)
if err ~= nil then
milla.send_message(err, "")
goto continue
end
local path, err = xmlpath.compile("//entry/title")
if err ~= nil then
milla.send_message(err, "")
goto continue
end
local iterator = path:iter(node)
for _, match in ipairs(iterator) do
table.insert(titles, match:string())
end
path, err = xmlpath.compile("//entry/author/name")
-- local path, err = xmlpath.compile("//entry/title")
if err ~= nil then
milla.send_message(err, "")
goto continue
end
iterator = path:iter(node)
for _, match in ipairs(iterator) do
table.insert(author_names, match:string())
end
path, err = xmlpath.compile("//entry/author/uri")
-- local path, err = xmlpath.compile("//entry/title")
if err ~= nil then
milla.send_message(err, "")
goto continue
end
iterator = path:iter(node)
for _, match in ipairs(iterator) do
table.insert(uris, match:string())
end
::continue::
end
for i = 1, #titles do
table.insert(rss_feed_list,
author_names[i] .. ": " .. titles[i] .. " -- " .. uris[i])
end
return rss_feed_list
end
local function rss_feed()
local config = get_config()
while true do
for _, v in pairs(get_rss_feed(config)) do
milla.send_message(v, config.channel)
sleep(config.period)
end
end
end
rss_feed()The example rss plugin, accepts a yaml file as input, reeds the provided rss feeds once, extracts the title, author name and link to the resource, sends the feed over to the #rssfeed irc channel and exits.
Also please note that this is just an example script. If you want milla to handle some rss feeds for you, you can use the builtin rss functionality.
Here's a list of the available function in milla's lua module:
milla.send_message(msg, target)milla.join_channel(channel,password)Please note that even if the channel doesn't have a password, it still will require the second argument and the second argument must be empty.
milla.part_channel(channel)milla.send_ollama_request(prompt)milla.send_gemini_request(prompt)milla.send_chatgpt_request(prompt)milla.send_or_request(prompt)milla.query_db(query)milla.register_cmd(script_path, cmd_name, function_name)milla.url_encode(str)Using register_cmd we can register a command that will be available to run like the built-in and customs commands.
Here's an example of how to use it:
local milla = require("milla")
local os = require("os")
local json = require("json")
-- setting the proxy value before loading the http module
-- this way, only this script will be using this proxy
os.setenv("ALL_PROXY", "socks5://172.17.0.1:9057")
local http = require("http")
-- this function should be global
-- one string arg that holds all args
-- should only return one string value
function milla_get_ip(arg)
local ip = arg
local response, err = http.request("GET", "http://ip-api.com/json/" .. ip)
if err ~= nil then print(err) end
local json_response, err = json.decode(response.body)
if err ~= nil then print(err) end
for k, v in pairs(json_response) do print(k, v) end
local result = ""
for key, value in pairs(json_response) do
result = result .. key .. ": " .. value .. " -- "
end
return result
end
milla.register_cmd("/plugins/ip.lua", "ip", "milla_get_ip")This will allow us to do:
terra: /ip 1.1.1.1And get this in response:
isp: Cloudflare, Inc -- query: 1.1.1.1 -- status: success -- regionName: Queensland -- lat: -27.4766 -- timezone: Australia/Brisbane -- region: QLD -- lon: 153.0166 -- country: Australia -- countryCode: AU -- city: South Brisbane --ip: 4101 -- org: APNIC and Cloudflare DNS Resolver project -- as: AS13335 Cloudflare, Inc. ---
Each lua plugin gets its own lua state and will run in a goroutine.
-
Lua plugins will not go through a proxy if they are not instructed to do so. If you are using the provided http module, you can set the proxy value before loading the http module as provided in the examples under
plugins. The module will read and set the following environment variables in the order given:ALL_PROXYHTTPS_PROXYHTTP_PROXYhttps_proxyhttp_proxy
httpandsocks5proxies are supported. unfortunately, thesocks5hproxy is not supported.
ALL_PROXY=socks5://172.17.0.1:9050More of milla's functionality will be available through milla's lua module over time.
The following libraries are loaded by milla by default:
- I end up with color escape sequences getting printed at the end of a line/begging of the next line. What gives?
This is happening because you have reached the message limit on irc which 512 for the event. This practically leaves around 390-400 character left for the message itself. Certain ircds allow for bigger sizes and certain clients might do. But most ircds don't send
linelento the clients. In a closed-loop situation where you control everything, as in, the ircd and all the clients(i.e. A private irc network), you can try to increase thelinelenfor the ircd and the client. Please note that the client in this case is girc. You irc client can have its own set of limits too. The 512 limit is hardcoded in girc. You can vendor the build or use the vendored dockerfile, change the hard limit and run milla with an increased limit. Needless to say, you can try to use achromaFormatterthat produces less characters which is basically not using truecolor orterminal16m.
Milla would not exist without the following projects: