[Bug]: SUPERMEMORY_EMBEDDING_PROVIDER and related env vars are ignored — embedding hardcoded to local

[DengHaitao-2022]

Bug Description

The self-hosted Supermemory server ignores the SUPERMEMORY_EMBEDDING_PROVIDER, SUPERMEMORY_EMBEDDING_MODEL, SUPERMEMORY_EMBEDDING_DIMENSIONS, and SUPERMEMORY_EMBEDDING_BASE_URL environment variables documented at https://supermemory.ai/docs/self-hosting/embeddings. The embedding provider is hardcoded to local with model Xenova/bge-base-v1.5 (768d).

Environment Variables Not Read

The binary only references SUPERMEMORY_EMBEDDING_RAM_LIMIT — none of the following documented variables are present in the compiled binary:

• SUPERMEMORY_EMBEDDING_PROVIDER
• SUPERMEMORY_EMBEDDING_MODEL
• SUPERMEMORY_EMBEDDING_DIMENSIONS
• SUPERMEMORY_EMBEDDING_BASE_URL

Proof (binary analysis)

# Only EMBEDDING_RAM_LIMIT is found
strings /usr/local/bin/supermemory-server | grep -o "SUPERMEMORY_EMBEDDING[A-Z_]*" | sort -u
# Output: SUPERMEMORY_EMBEDDING_RAM_LIMIT

# Default provider hardcoded
strings /usr/local/bin/supermemory-server | grep "var y90"
# Output: var y90="local"

Steps to Reproduce

1. Set environment variables in /etc/supermemory/supermemory.env:

SUPERMEMORY_EMBEDDING_PROVIDER=openai
SUPERMEMORY_EMBEDDING_MODEL=BAAI/bge-m3
SUPERMEMORY_EMBEDDING_DIMENSIONS=1024
SUPERMEMORY_EMBEDDING_BASE_URL=http://10.10.0.2:3004/v1

2. Restart the service
3. Check logs — still shows * local embeddings Xenova/bge-base-v1.5
4. Verify env vars are loaded in process

Expected Behavior

The server should use the configured embedding provider (e.g., OpenAI-compatible endpoint with BAAI/bge-m3).

Actual Behavior

The server always uses local Xenova/bge-base-v1.5 (768d) regardless of environment variable configuration.

Affected Versions

• v0.0.6 (stable)
• v0.0.7-rc.2 (pre-release)

Impact

Users with non-English content cannot use the self-hosted server effectively because the default local model is English-only. The documented workaround (switching to remote embedding or local bge-m3) does not work.

Additional Context

• Platform: Ubuntu 24.04, x86_64
• Installation: systemd service
• The OpenAI-compatible endpoint works correctly when tested directly via curl
• The LLM configuration (OPENAI_BASE_URL, OPENAI_MODEL) works fine — only embedding is affected

[linear-code]

ENG-1104

[100-eyes]

Independent confirmation against the official Linux x64 server-v0.0.7-rc.2 release asset:

• Release commit: 816b85d
• Asset SHA-256: f56e7abc1d2d838d75c47827d0a823672d3707f2e0c26c4472d15805080d9a8c
• Platform: Linux x86_64
• Test used a completely fresh Supermemory state directory and isolated listener

Configuration:

SUPERMEMORY_EMBEDDING_PROVIDER=local
SUPERMEMORY_EMBEDDING_MODEL=Xenova/bge-m3
SUPERMEMORY_EMBEDDING_DIMENSIONS=1024
SUPERMEMORY_LOCAL_EMBEDDING_POOL_SIZE=1
SUPERMEMORY_LOCAL_EMBEDDING_WASM_THREADS=1
SUPERMEMORY_LOCAL_EMBEDDING_BATCH_SIZE=8

We verified through /proc/<pid>/environ that all six values reached the running process. Nevertheless, first boot reported:

* local embeddings  Xenova/bge-base-en-v1.5 · first-run download
    pool: 1 worker · 1 thread each

The binary then downloaded the 106 MB default English model.

Additional release-artifact evidence:

Literal in official rc.2 Linux binary	Count
Xenova/bge-base-en-v1.5	1
Xenova/bge-m3	0
SUPERMEMORY_EMBEDDING_PROVIDER	0
SUPERMEMORY_EMBEDDING_MODEL	0
SUPERMEMORY_EMBEDDING_DIMENSIONS	0

The local-worker tuning variables are present in the same binary. This appears to be a release-build capability mismatch: the documentation at the rc.2 tag describes pluggable embeddings, while the published binary includes the tuning knobs but not provider/model/dimension selection.

We also tested first-run setup in interactive TTY mode:

1. With the LLM provider supplied through environment variables, setup skipped the embedding picker.
2. Without LLM environment variables, the wizard asked for LLM provider/key/base URL/model and then completed without an embedding picker.

This was reproduced before any benchmark ingestion and without Hermes, our privacy adapter, or Tinfoil in the request path.

Practical multilingual impact from a separate controlled 12-fact German test on the default model:

• Raw German queries: Hit@3 6/12
• German paraphrases: Hit@3 7/12
• English translations of the same queries: Hit@3 12/12

Happy to retest a corrected Linux x64 release asset with Xenova/bge-m3.