InnieMe

InnieMe is a chat bot that answers questions from your own documents. Point it at a directory of files, connect it to a chat channel, and it responds to mentions with context-aware answers backed by a vector search over your knowledge base — using the LLM provider of your choice (OpenAI, Anthropic, …) via PydanticAI.

It currently supports Discord and Slack. It's built for teams that keep answering the same questions: instead of repeating yourself (DRY), let InnieMe field the routine ones from your docs and keep each interaction in its own thread.

Features

• Multi-platform — run the same knowledge bot on Discord or Slack from one codebase and CLI.
• Document-grounded answers — scans and vectorizes a documents directory and uses similarity search to ground every response.
• Pluggable models — choose your embedding backend (openai, huggingface, or fake for testing) and any PydanticAI LLM (e.g. openai:gpt-5.6-terra, anthropic:claude-sonnet-5).
• Threaded conversations — each mention spins up a thread and the bot follows it for context.
• Multi-topic — define multiple topics, each with its own system prompt, documents, and channels, owned by one or more admins who own the documents.

How it works

1. On startup the bot reads its config, vectorizes the documents for each configured topic, and connects to the chat platform.
2. When mentioned in a watched channel, it retrieves the most relevant document chunks, builds a prompt (topic role + context + conversation history), and replies in a thread. Each chunk is labelled with the file it came from, so the model can attribute an answer to a source document.

Prerequisites

• Python 3.13+
• A bot for your platform (Discord bot token, or Slack bot + app tokens) and the IDs of the server/channel(s) it should watch
• An API key for your chosen LLM provider (and for embeddings, if using OpenAI embeddings)

discord_config.example.yaml (Discord) and slack_config.example.yaml (Slack) include step-by-step instructions for obtaining each token/ID.

Installation

git clone https://github.com/wolfdancer/innieme.git
cd innieme
# (recommended) create a virtual environment
python -m venv .venv &&source .venv/bin/activate
# install the package and its dependencies
pip install -e .

Configuration

Copy the example config for your platform and fill in your values:

cp discord_config.example.yaml discord_config.yaml # Discord
cp slack_config.example.yaml slack_config.yaml # Slack

Common fields (both platforms):

Field	Default	Description
embedding_model	—	"openai", "huggingface", or "fake" (use fake in tests to avoid API calls)
embeddings_model_name	per backend	Embedding model name. Unset means the backend's default: text-embedding-3-small (OpenAI) or all-MiniLM-L6-v2 (HuggingFace)
embeddings_api_key	—	API key for the embedding model (required for openai)
llm_model	openai:gpt-5.6-terra	PydanticAI model string, e.g. "openai:gpt-5.6-terra" or "anthropic:claude-sonnet-5"
llm_api_key	—	API key for the LLM provider
cache_dir	<docs_dir>/.cache/langchain	Where downloaded embedding models are cached. Only used by the huggingface backend; supports ~
retrieval_top_k	5	Maximum document chunks sent to the model as context per query
retrieval_score_threshold	unset	Optional relevance floor (0–1). Drops weak matches instead of padding context out to retrieval_top_k
outies	—	List of admins, each with one or more topics

Per-topic fields, inside each entry of a topic list:

Field	Default	Description
name	—	Topic name
role	—	The topic's system prompt
docs_dir	—	Directory of documents to ingest for this topic
docs_exclude	["CLAUDE.md"]	Filename patterns to skip when scanning this topic's docs_dir. Set to [] to scan everything
channels	—	Channels where this topic answers

Tuning retrieval

retrieval_top_k caps how much document context each answer is built from. Raising it improves recall on questions that span several documents, at the cost of more input tokens — and of more irrelevant context, which degrades answer quality more often than it helps.

retrieval_score_threshold makes the count adaptive: chunks are still capped at retrieval_top_k, but any scoring below the floor are dropped, so a narrow question returns only the two or three chunks that actually matter. It is unset by default, which is the safe choice — setting it too high makes the bot report that something is absent when it is present. Pick a value by measuring rather than guessing: compare the scores for questions you know your documents answer against questions you know they do not, and choose a value between the two ranges. If the ranges overlap, no threshold will separate them and it should stay off.

Chroma collections use cosine distance, which is the appropriate metric for text embeddings and keeps relevance scores in a usable 0–1 range.

Excluding files from the knowledge base

docs_exclude is set per topic, next to that topic's docs_dir — different document sets have different non-content files, so one global list would be wrong for most of them. Each pattern is matched against both the filename and the path relative to docs_dir, so CLAUDE.md skips that file at any depth while archive/* skips a subdirectory.

It defaults to ["CLAUDE.md"]. Agent instruction files are not subject-matter content, and ingesting them is actively harmful: a retrieved chunk of instructions reads to the model as directions to follow, so rules written for one task ("always produce a next action, even if it's a guess") can override the answering prompt's rules ("never invent a next step"). Set docs_exclude to [] to scan everything.

The startup message reports how many files were excluded, and the logs name each one. The count goes to the channel but the names do not: a file is often excluded precisely because the people in that channel should not know about it, while whoever configured the bot can read the logs.

Platform-specific fields:

• Discord: discord_token; outie_id/guild_id/channel_id are numeric IDs.
• Slack: slack_bot_token + slack_app_token (Socket Mode); outie_id (U…) and channel_id (C…) are strings.

Place the documents you want the bot to learn from in the docs_dir configured for each topic.

Running

Use the unified CLI and pick a platform. It loads the config from the current working directory by default (run from the project root), or pass -c to point elsewhere:

innieme discord # uses ./discord_config.yaml
innieme slack # uses ./slack_config.yaml
innieme discord -c custom_config.yaml

Logging is controlled by environment variables: LOG_LEVEL (global, default INFO) and INNIEME_LOG_LEVEL (this package, default INFO).

Slack commands

Ask a question by mentioning the bot, or by replying in a thread it is already following. The bot also understands three commands, given the same way:

Command	Who can use it	What it does
@bot hello	anyone	Posts the introduction card. Works in any channel, even one with no topic configured, so it doubles as an "is this thing running?" check.
@bot rescan	the topic's outie	Re-reads and re-vectorizes the topic's docs_dir. Use it after editing your documents — there is no need to restart. If the scan fails, the previous index keeps serving answers.
@bot quit	the topic's outie	Shuts the bot down, process included.

The whole message has to be the command, so @bot rescan runs a rescan while @bot should we rescan the notes? is answered as a question.

These are mentions, not slash commands, deliberately: a slash command must be declared in your Slack app's configuration as well as in the code, so it cannot work on a fresh install without extra setup, whereas a mention works as soon as the bot is running.

/approve (approve a generated summary into the knowledge base) is the one remaining slash command, so it does need declaring under Features → Slash Commands in your Slack app to be reachable.

Upgrading:/quit and /hello used to be slash commands and are now the mentions above. If you declared either in your Slack app configuration, delete it there — otherwise Slack keeps offering a command the bot no longer handles.

Docker

docker build -t innieme .# Discord (default command)
docker run -v "$(pwd)/discord_config.yaml:/app/discord_config.yaml" -v "$(pwd)/data:/app/data" innieme
# Slack — pass the platform as the command
docker run -v "$(pwd)/slack_config.yaml:/app/slack_config.yaml" -v "$(pwd)/data:/app/data" innieme slack

Development

# install dev dependencies
pip install -e .&& pip install -r requirements-dev.txt
# run the test suite
pytest
# format, sort imports, and lint
black src/ tests/
isort src/ tests/
flake8 .

License

See LICENSE.