embedcrate/embed.env.example at main · hwdsl2/embedcrate · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
# Note: All the variables to this image are optional.
# See README for more information.
# To use, copy this file to 'embed.env' and uncomment/set your values.

# Embedding model to use (default: BAAI/bge-small-en-v1.5)
# Pass any HuggingFace model ID supported by Hugging Face TEI.
# Recommended models:
#   BAAI/bge-small-en-v1.5              ~130 MB   Fastest; English
#   BAAI/bge-base-en-v1.5               ~440 MB   Good balance; English
#   BAAI/bge-large-en-v1.5              ~1.3 GB   High accuracy; English
#   BAAI/bge-m3                          ~570 MB   Multilingual
#   nomic-ai/nomic-embed-text-v1.5       ~550 MB   Strong multilingual; long context
#   sentence-transformers/all-MiniLM-L6-v2  ~90 MB   Very small; fast
# - Models are downloaded from HuggingFace on first start and cached in the
#   /var/lib/embeddings Docker volume.
# EMBED_MODEL=BAAI/bge-small-en-v1.5

# HTTP port for the embeddings API (default: 8000)
# EMBED_PORT=8000

# Optional API key to require Bearer token authentication on all requests.
# - Fresh installs with a mounted /var/lib/embeddings volume auto-generate a key.
# - Existing installs with no key remain open for backward compatibility.
# - When set, every request must include: -H "Authorization: Bearer <key>"
# - Set explicitly empty (EMBED_API_KEY=) to disable authentication.
# - Use 'docker exec <container> embed_manage --showkey' to display the key.
# - Use 'docker exec <container> embed_manage --getkey' for scripts.
# - Keep this secret if you expose the server to the internet.
# EMBED_API_KEY=

# HuggingFace Hub token for accessing private or gated models (default: unset)
# - Only needed for private or gated models. Public models (BAAI/bge-*, nomic-*)
#   do not require a token.
# EMBED_HF_TOKEN=

# Local-only mode: disable all HuggingFace model downloads (default: unset)
# - When set to any non-empty value (e.g. true), the server will only use models
#   already present in the /var/lib/embeddings cache volume and will not attempt
#   to contact HuggingFace. Useful for air-gapped or offline deployments.
# - Leave unset (default) to allow automatic model downloads on first start.
# EMBED_LOCAL_ONLY=

# Embeddings server toggle (default: true)
# Set to "false" to disable the embeddings process. Useful for rerank-only mode.
# EMBED_ENABLED=true

# Anonymous aggregate usage counts (default: enabled)
# Set to 1 to disable.
# EMBED_DISABLE_USAGE_COUNTS=1

# ============================================================
# Reranking (optional — disabled by default)
# ============================================================
# Enable the reranking server. When set to "true", a second TEI process starts
# serving a cross-encoder model on RERANK_PORT. The reranker provides a
# POST /rerank endpoint for re-scoring retrieved documents by relevance.
# RERANK_ENABLED=

# Cross-encoder model for reranking (default: BAAI/bge-reranker-v2-m3)
# Recommended models:
#   BAAI/bge-reranker-v2-m3              ~560 MB   Multilingual; strong accuracy — default
#   BAAI/bge-reranker-base               ~440 MB   English; good balance
#   BAAI/bge-reranker-large              ~1.3 GB   English; highest accuracy
#   cross-encoder/ms-marco-MiniLM-L6-v2   ~90 MB   Very small; fast; English
# RERANK_MODEL=BAAI/bge-reranker-v2-m3

# HTTP port for the reranker API (default: 8001, or 8000 if embeddings is disabled)
# RERANK_PORT=8001

# Optional API key for the reranker. If unset, falls back to EMBED_API_KEY.
# Set explicitly empty (RERANK_API_KEY=) to disable reranker authentication.
# RERANK_API_KEY=