Getting Started

Prerequisites

• Miniconda
• A working Kalisio development stack

Install Miniconda

Install Miniconda with the official installer:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O /tmp/miniconda_installer.sh
bash /tmp/miniconda_installer.sh -b -p ~/miniconda3
~/miniconda3/bin/conda init bash

Reload your shell:

source ~/.bashrc

Check the installation:

conda --version

Install Project Dependencies

The project uses Python 3.11, managed through environment.yml.

Create the conda environment and install all dependencies:

conda env create -f environment.yml

Activate the environment:

conda activate knowledge

Environment Variables

Shared (API + ingestion job)

Variable	Required	Default	Description
QDRANT_URL	yes	-	URL of the Qdrant instance (http://localhost:6333)
QDRANT_COLLECTION_CODE	yes	-	Collection for source-code chunks (ingestion)
QDRANT_COLLECTION_METADATA	yes	-	Collection for metadata chunks (ingestion)
EMBEDDING_MODEL	yes	-	HuggingFace model identifier - see tip below
EMBEDDING_BATCH_SIZE	no	8	Chunks embedded per batch during ingestion
LOG_LEVEL	no	INFO	Log verbosity (DEBUG, INFO, WARNING, …)

API service

Variable	Required	Default	Description
LLM_API_KEY	yes	-	API key for the LLM provider
LLM_MODEL	yes	-	Chat model identifier (e.g. gpt-4o-mini)
LLM_ENDPOINT	yes	-	Base URL of any OpenAI-compatible API
APP_SECRET	yes*	-	Secret to sign and verify JWT tokens (* required when auth is on)
KNOWLEDGE_AUTH_ENABLED	no	true	Set to false to disable JWT auth
JWT_AUDIENCE	no	kalisio	Expected aud claim for incoming tokens
JWT_ISSUER	no	kalisio	Expected iss claim for incoming tokens
JWT_ALGORITHM	no	HS256	Algorithm used to verify JWT signatures
TOP_K	no	6	Number of chunks returned per retrieval
MAX_CONTEXT_CHARS	no	14000	Max characters of context fed to the LLM
MAX_ANSWER_TOKENS	no	1024	Max tokens the LLM may generate per answer
HOST	no	127.0.0.1	Host the API binds to
PORT	no	8187	Port the API listens on

Ingestion job

Variable	Required	Default	Description
KALISIO_GITHUB_TOKEN	yes	—	Token k-clone clones the GitHub repositories with. Without it the tooling falls back to ssh, which needs a key in the container
GITLAB_IRSN_TOKEN	no	—	Same, for an IRSN workspace hosted on GitLab
KLI_ORGANIZATION	no	kalisio	Organisation passed to k-clone (kalisio, irsn, airbus)
KLI_WORKSPACE	no	apps	kli workspace to clone and index
INDEXED_REPOSITORIES	no	—	Comma-separated repository names to index; empty means all of them
COMMIT_HISTORY_MAX_AGE_DAYS	no	180	Age window of the commit history kept per file
COMMIT_HISTORY_MIN_COMMITS	no	3	Commits kept however old, so a stable file keeps a history
COMMIT_HISTORY_DEPTH	no	0	Cap on the commits kept per file; 0 means no cap

Choosing an embedding model

EMBEDDING_MODEL accepts any SentenceTransformers model identifier from HuggingFace(Qwen/Qwen3-Embedding-0.6B, Qwen/Qwen3-Embedding-4B, nomic-ai/nomic-embed-text-v1.5)

OpenAI models for the LLM

LLM_ENDPOINT follows the OpenAI API format, so any OpenAI-compatible provider works without code changes: OpenAI (gpt-4o, gpt-4o-mini, gpt-4.1), Mistral, Ollama (local), and others.

Run the Project

Start the API:

python -m api.bin

Run the ingestion job:

python -m ingestion