.env.example

# Example environment file for TalkPipe Vault
# Copy this to .env and customize for your deployment

# Web server configuration
VAULT_HOST=0.0.0.0
VAULT_PORT=8002

# Host folder mounted read-only at /documents in the container (compose only);
# browse to it in the folder picker when indexing. Must exist on the host.
# VAULT_DOCUMENTS_DIR=~/Documents

# Path fences enforced by the web app itself (the container image defaults to
# these values; outside a container both are unset, i.e. unrestricted):
# - TALKPIPE_VAULT_ROOT: vaults may only be created/opened/deleted under this
#   single directory.
# - TALKPIPE_DOCUMENT_ROOTS: folder browsing and indexing are limited to these
#   directories (separate several with ":", the os.pathsep character).
# Set a variable to empty to lift that restriction.
# TALKPIPE_VAULT_ROOT=/app/data
# TALKPIPE_DOCUMENT_ROOTS=/documents

# LanceDB directory path (same semantics as makevectordatabase --path);
# Whoosh index is stored under VAULT_PATH/fulltext_vault
VAULT_PATH=/app/data/vault

# Where web-interface settings (recent vaults, model choices) are persisted
TALKPIPE_VAULT_HOME=/app/data/vault-home

# Model configuration. Defaults: embeddings use model2vec
# (minishlab/potion-retrieval-32M), which runs fully in-process — no server
# or API key; chat uses ollama/mistral-small. Uncomment to override:
# TALKPIPE_EMBEDDING_SOURCE=model2vec
# TALKPIPE_EMBEDDING_MODEL=minishlab/potion-retrieval-32M
# TALKPIPE_CHAT_SOURCE=ollama
# TALKPIPE_CHAT_MODEL=mistral-small

# Firewalled / air-gapped deployments (model2vec embeddings):
# The embedding model (default minishlab/potion-retrieval-32M) is downloaded
# from Hugging Face on first use into the cache in the data volume, so it is
# fetched only once. When HF_HUB_OFFLINE is unset, the entrypoint probes
# huggingface.co at startup and automatically switches to offline (cache-only)
# loads if it is unreachable, so nothing hangs on connection timeouts; if the
# model is not cached yet either, the server still starts (without a vault)
# and reports the problem on the Settings page. Override the auto-detection
# only if needed:
# HF_HUB_OFFLINE=1   # force offline: never contact Hugging Face
# HF_HUB_OFFLINE=0   # force online: always check Hugging Face (loads can
#                    # stall when huggingface.co is slow or blocked)
# For a host that can never reach Hugging Face, make the model available
# locally by either:
#   (a) pointing the embedding model at a downloaded directory (most reliable —
#       no Hugging Face lookup at all), e.g.
#       TALKPIPE_EMBEDDING_MODEL=/app/data/models/potion-retrieval-32M
#   (b) pre-populating the Hugging Face cache in the data volume from a
#       connected machine.
# HF_HOME is already set to /app/data/hf-cache in the Containerfile, so
# downloaded models persist in the data volume; override only to relocate it.
# HF_HOME=/app/data/hf-cache

# Optional: provider connection settings
# OPENAI_API_KEY=sk-your-key-here
# TalkPipe reads the Ollama server URL from TALKPIPE_OLLAMA_SERVER_URL.
# Use http://host.containers.internal:11434 when Ollama runs on the container host.
# Use a network hostname or IP when Ollama runs elsewhere, for example:
TALKPIPE_OLLAMA_SERVER_URL=http://your-ollama-host:11434