EDR-Constrained Windows (CrowdStrike): Remote Embeddings

On a locked-down corporate Windows machine — typically a VDI running CrowdStrike Falcon or a similar EDR/AV — the local ONNX embedding backend does not work. The local-embed / local-embed-dynamic features make ort/fastembed download an unsigned onnxruntime.dll, and the EDR’s quarantine-on-write deletes it as malware seconds after it lands on disk — the same heuristic that flags freshly-built unsigned executables. Semantic search then silently degrades to the lexical (SQL LIKE) keyword fallback.

You usually cannot fix this by whitelisting — AV exclusions are out of your control on a managed VDI. The fix is to keep ONNX off the box entirely and run embeddings against a remote endpoint.

Fix: remote embeddings, no ONNX on the box

codescout’s Embedder is backend-agnostic. Point it at any OpenAI-compatible embedding endpoint — a corporate internal embedding API, Ollama, TEI, etc. — and the local ONNX runtime is never compiled or downloaded.

1. Build without the local ONNX feature

The default feature set already excludes local ONNX:

cargo build --release        # default features = remote-embed, http, librarian

Do not pass --features local-embed or --features local-embed-dynamic. After building, confirm there is no onnxruntime.dll anywhere under target/ and that ort / fastembed were not compiled. Now the EDR has nothing to quarantine.

2. Point at the embedding endpoint

In <project>/.codescout/project.toml:

[embeddings]
model   = "your-model-name"                # the bare model name the API expects
url     = "https://embed.corp.example/v1"  # OpenAI-compatible; /v1/embeddings is appended
api_key = "..."                            # or set EMBED_API_KEY in the environment

Setting url takes priority over the model prefix, so resolution routes straight to the remote backend (RemoteEmbedder) — any local: prefix on model is stripped. Equivalent environment variables: CODESCOUT_EMBED_URL, CODESCOUT_EMBED_MODEL, EMBED_API_KEY.

HTTPS is required when api_key is set. codescout refuses to send an API key over plaintext HTTP. Loopback hosts (localhost, 127.0.0.1, [::1]) are exempt — that carve-out is for local Ollama / llama.cpp, not a network endpoint. Use an https:// URL for a corporate API.

3. Reindex (do not skip)

The default local model (AllMiniLML6V2Q) is 384-dimensional; your remote model almost certainly emits a different dimension (often 768 / 1024 / 1536). The vector index is dimension-specific, so a stale index built with the old dimension produces wrong or empty results. Rebuild the index after switching embedders (the index tool / your normal reindex path) — config change first, reindex second.

4. Verify

• semantic_search returns results.
• No onnxruntime.dll exists under target/.
• The EDR quarantines nothing during build or run.

Daemon-free: the lite stack (no Qdrant, no sparse, no reranker)

Canonical setup and the two-stack overview now live in The Lite Stack — it is the default build, not just the EDR case. This section keeps the EDR-specific framing.

If the VDI cannot run Docker or Qdrant at all — the common locked-down case — use the lite stack. Code search and memory run entirely in-process on sqlite-vec (a statically-linked vec0 table — no foreign DLL for the EDR to quarantine, unlike onnxruntime.dll), with dense embeddings from your remote endpoint. No Qdrant, no sparse SPLADE server, no cross-encoder reranker.

# or copy .env.lite from the repo root and `source` it
export CODESCOUT_VECTOR_BACKEND=sqlite-vec
export CODESCOUT_EMBEDDER_URL=https://embed.corp.example/v1
export CODESCOUT_EMBEDDER_MODEL_NAME=your-model-name
export CODESCOUT_MODEL_DIM=768
export EMBED_API_KEY=...   # sent only over HTTPS (loopback exempt)

Per-project indexes live under CODESCOUT_SQLITE_DIR (default <home>/.codescout/embeddings). Then run index(action='build') once per project. The trade-off versus the full server stack is dense-only ranking — no sparse exact-token leg, no rerank — so exact-identifier recall is weaker; pair it with a strong code-embedding model on the endpoint. Design + rationale: docs/plans/2026-06-16-two-stack-retrieval-lite.md.

Why not just sign the DLL?

Authenticode-signing the ONNX runtime would beat the quarantine-on-write of the unsigned file, but CrowdStrike Falcon is behavior-first: ML runtimes that allocate PAGE_EXECUTE_READWRITE memory or use dynamic LoadLibrary can still trip the behavioral engine even when signed. Remote embeddings remove the runtime from the box altogether, so there is nothing to sign, load, or quarantine.

If no endpoint is reachable (air-gapped)

If the machine is fully air-gapped with no reachable embedding endpoint:

• Lexical fallback — with no embedder configured, semantic_search falls back to SQL LIKE keyword search. Functional, but not semantic.
• Pure-Rust embeddings (candle) — a future candle-based local backend would run the model in-process with no foreign DLL (pure-Rust gemm, no PAGE_EXECUTE_READWRITE codegen), compiling into the large codescout.exe that already survives the EDR’s tiny-PE heuristic. Not yet implemented — tracked as a candidate backend.