js-langchain: remove native hnswlib-node dependency (Windows install fails without C++ toolchain)

[l2ysho]

Background

The js-langchain template depends on hnswlib-node (via HNSWLib from @langchain/community/vectorstores/hnswlib), which is a native C++ addon compiled with node-gyp at install time.

On Windows without a node-gyp–compatible Visual Studio toolchain, npm install fails:

npm error path .../node_modules/hnswlib-node
npm error command ... node-gyp rebuild
gyp ERR! find VS could not find a version of Visual Studio 2017 or newer to use
gyp ERR! stack Error: Could not find any Visual Studio installation to use

This surfaced in CI when the windows-latest runner image moved to Visual Studio 18 (2026), which the bundled node-gyp can't detect. It also affects real users on Windows who don't have the "Desktop development with C++" workload installed.

CI mitigation already in place

The LLM-AI-node test workflow is pinned to windows-2022 (which ships VS 2022 that node-gyp supports) as a stopgap. That only fixes CI — it does not fix the install for end users, and windows-2022 is on GitHub's deprecation path.

Proposed fix — drop the native dependency

Replace the native HNSWLib vector store with a pure-JS store so the template installs on any platform with no C++ toolchain.

• Swap HNSWLib for a pure-JS vector store (e.g. MemoryVectorStore from langchain/vectorstores/memory), or another store that doesn't require a native build.
• Remove hnswlib-node from package.json.
• Rework src/vector_index_cache.js: it currently tars HNSWLib's on-disk index directory into the vector-index-cache key-value store and untars it back. MemoryVectorStore has no save(path)/load(path) on-disk format, so the index caching must either be re-implemented (e.g. serialize the in-memory vectors to JSON in the KV store) or dropped.

Files involved

• templates/js-langchain/src/main.js — HNSWLib.fromDocuments / vectorStore.save / HNSWLib.load
• templates/js-langchain/src/vector_index_cache.js — persistence/caching layer
• templates/js-langchain/package.json — drop hnswlib-node

Considerations / open questions

• This is user-facing example code. Moving to an in-memory store changes what the template demonstrates and likely degrades or removes the "cache the vector index between runs" feature. We should decide whether to preserve caching (via JSON serialization) or accept its removal.
• Persistent-but-pure-JS options in LangChain JS are limited (FAISS is also native; Chroma/PGVector need a server), which is why MemoryVectorStore + custom JSON caching is the most likely path.

Acceptance criteria

• npm install for js-langchain succeeds on Windows with no Visual Studio / C++ build tools.
• Template runs end-to-end on ubuntu and windows-latest.
• Once merged, revert the windows-2022 pin in .github/workflows/test_llm_ai_node_templates.yaml back to windows-latest.

[wiaikit]

This is a good case for making the install path explicit for Windows users.

For native modules on Windows, there are really two different outcomes:

1. a prebuilt binary exists, so users do not need a local C++ toolchain;
2. no matching prebuild exists, so npm falls back to node-gyp rebuild and users need Python + Visual Studio Build Tools / Desktop development with C++.

For a user-facing template, I would avoid making branch 2 the default path if the native dependency is optional. If it must stay, the README should probably include a short Windows diagnostic before telling users to install the full toolchain:

node -v
npm -v
npm config get python
where.exe cl
where.exe msbuild

I maintain a free Windows setup diagnostic/checklist for this exact node-gyp / Build Tools class of failures:
https://pay.wiaikit.com/free?lang=en&from=apify-813