All settings are managed through the .env file (environment variables can override it).

cp .env.example .env

Full reference

# ── LLM ──
LLM_API_BASE=https://api.deepseek.com/v1
LLM_API_KEY=sk-xxx
LLM_MODEL=deepseek-chat

# ── Embedding ──
USE_LOCAL_EMBEDDING=false
EMBEDDING_API_BASE=https://api.deepseek.com/v1
EMBEDDING_API_KEY=sk-xxx
EMBEDDING_MODEL=text-embedding-3-small
LOCAL_EMBEDDING_MODEL=BAAI/bge-small-zh-v1.5    # CPU-friendly, ~100MB

# ── Retriever ──
RETRIEVER_MODE=simple          # simple / multi_query
SIMILARITY_TOP_K=3
MULTI_QUERY_COUNT=3

# ── Multimodal ──
IMAGE_EMBEDDING_MODE=disabled  # disabled / clip / vlm_describe
CLIP_MODEL=openai/clip-vit-base-patch32
VLM_MODEL=                     # empty = reuse LLM_MODEL
USE_MULTIMODAL_LLM=false       # pass retrieved images to multimodal LLM when answering
IMAGE_SIMILARITY_TOP_K=2

# ── Memory ──
MEMORY_TOKEN_LIMIT=30000
CHAT_HISTORY_TOKEN_RATIO=0.7

# ── Data Profile ──
DATA_PROFILE=default           # switch knowledge-base profile

Data Profile

Use DATA_PROFILE to switch between knowledge bases. Data and indexes are fully isolated per profile:

DATA_PROFILE=default python main.py     # default profile
DATA_PROFILE=2wiki python main.py       # switch to 2wiki dataset

Corresponding paths:

• Data: data/profiles/{DATA_PROFILE}/
• Indexes: storage/{DATA_PROFILE}/

Multimodal settings

Variable	Description
IMAGE_EMBEDDING_MODE	disabled (default) / clip / vlm_describe
CLIP_MODEL	CLIP model name (used in clip mode)
VLM_MODEL	VLM model for vlm_describe (empty = reuse LLM_MODEL)
USE_MULTIMODAL_LLM	Whether to pass retrieved images to a multimodal LLM when answering
IMAGE_SIMILARITY_TOP_K	Number of image hits to return (clip mode)

Install llama-index-embeddings-clip for clip mode and llama-index-multi-modal-llms-openai for vlm_describe mode.

How it works

DataMind loads configuration with Pydantic Settings:

1. Read the .env file
2. Environment variables override values from .env
3. Unset variables fall back to built-in defaults

Temporary overrides:

LLM_MODEL=gpt-4o RETRIEVER_MODE=multi_query python main.py

Paths

Property	Default	Description
data_dir	data/profiles/{DATA_PROFILE}/	Profile-scoped data directory
storage_dir	storage/{DATA_PROFILE}/	Index persistence directory
skills_dir	data/skills/	Skill documents (shared across profiles)
bench_dir	data/bench/	Question sets (shared across profiles)