[BUG] OpenAI embedding provider silently falls back to random vectors while reporting success

[gudzenkov]

Summary

When built with OpenAI embedding support and run with CODEGRAPH_EMBEDDING_PROVIDER=openai, indexing prints a critical fallback to random/hash embeddings, but the final indexing summary still reports successful OpenAI embeddings.

This makes a degraded index look valid.

Related

• Related to Issue with js/ts project, files not indexed #43, which also reports random embedding fallback, but in a mixed TS/JS indexing report rather than the OpenAI provider path specifically.
• No matching PR found during search.

Clean Reproduction

Fresh clone:

git clone --depth 1 https://github.com/Jakedismo/codegraph-rust codegraph-rust-clean-openai
cd codegraph-rust-clean-openai
git rev-parse HEAD
# ce5bf27 Merge pull request #59 from Jakedismo/wip-indexing-analyzers-doc

Build:

cargo build --release -p codegraph-mcp-server --bin codegraph \
  --no-default-features \
  --features ai-enhanced,embeddings-openai,codegraph-ai/openai-llm

Index command:

CODEGRAPH_SURREALDB_URL=ws://127.0.0.1:51068 \
CODEGRAPH_SURREALDB_NAMESPACE=codegraph \
CODEGRAPH_SURREALDB_DATABASE=main \
CODEGRAPH_SURREALDB_USERNAME=root \
CODEGRAPH_SURREALDB_PASSWORD=root \
CODEGRAPH_EMBEDDING_PROVIDER=openai \
CODEGRAPH_EMBEDDING_MODEL=text-embedding-3-small \
CODEGRAPH_EMBEDDING_DIMENSION=1536 \
CODEGRAPH_LLM_PROVIDER=openai \
CODEGRAPH_MODEL=gpt-5.4-mini \
CODEGRAPH_CONTEXT_WINDOW=200000 \
RUST_LOG=info \
target/release/codegraph --config /path/to/codegraph-rust.toml \
  index src --recursive --index-tier fast --force --batch-size 64 --max-concurrent 2

Observed during indexing:

CRITICAL: Falling back to random hash-based embeddings

Observed final summary:

Embeddings: chunks 2157 | dim 1536 | provider openai

Database state:

chunks: 2157
embedding_1536 populated: 2157

Expected

One of:

• Real OpenAI embeddings are generated.
• Indexing fails hard with an actionable OpenAI provider error.
• The final summary clearly reports fallback/random embeddings and marks the index unsuitable for semantic retrieval.

Actual

The index completes and reports OpenAI embeddings even though the log says random hash-based embeddings were used.

Impact

Downstream semantic search depends on embedding quality. A silent random fallback can make retrieval quality invalid while the CLI summary says the OpenAI embedding path succeeded.