- AI tagging through any OpenAI-compatible API: DeepSeek by default, or
OpenRouter, OpenAI, or a local server (Ollama, LM Studio). Settings → AI
with presets, prices and "Save & test connection"; the key never leaves
the server except to the provider.
- Runs over untagged models, all models, or one model from its page: live
progress, tokens and cost, stop/resume (retrying failures), an optional
spending limit, and paused rather than resumed after a restart. Adapts to
models without JSON mode or image input; stops on a rejected key or no
credit. Tags are only added; descriptions filled only where empty unless
asked.
- Search by meaning: local bge-small-en-v1.5 embeddings (transformers.js),
indexed in the background whenever a model's text changes; Words/Meaning
switch on the search box. Stored as REAL[] in Postgres, no pgvector needed.
- Automatic tags: strict whole-word rules (part types, mechanisms, materials,
techniques) over names, file names and archive contents, plus your own
word -> tag rules; on/off (optionally removing its tags) and re-apply.
- Tags remember where they came from (user / rule / ai); edits keep that,
rule and AI passes never touch your own tags, and the model page marks them.
- Docker: ships the embedding model and only Linux onnxruntime binaries.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Archives
- ZIP, 7Z, RAR and TAR/.tar.gz uploads are indexed in the background into
archive_entries (7-Zip compiled to WASM in a worker thread; node-tar for
tar). macOS/Windows clutter and paths that could escape the extraction
folder are skipped.
- Models inside archives are extracted once per archive to hash and
geometry-fingerprint them, so duplicates are found inside packs too.
Pictures, PDFs and readmes are listed but never matched.
- New routes list an archive's entries and serve/convert a single entry,
extracted on first use and cached under extracted/.
Viewer
- Model detail lists what's inside each archive with view/download links;
opening an entry adds a viewer tab.
- New 3D formats: FBX (with sibling textures), OBJ with its MTL and
textures, STEP/STP via occt-import-js, and Rhino 3DM via rhino3dm — the
WASM readers are served from node_modules, loaded on first use.
- Pictures, PDFs and text files (jpg/png/gif/webp/bmp/svg, pdf, txt/md)
can be uploaded and preview in the viewer; a model without an embedded
3MF preview uses its first picture as the card thumbnail.
- Inline file responses send nosniff, and SVGs are sandboxed.
- GLB conversions write to a temp file and rename, so a half-written
preview is never served; a later tab can't be overdrawn by a slow load.
- Better 3MF card image: the project thumbnail or plate 1 render, not a
slicer pick map.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
trimesh's 3MF loader re-reads a whole production-extension part file for
every component that points into it. A Bambu Studio project with 884
components across four ~50 MB part files grew past 3 GB within seconds
and failed both the preview conversion and the geometry fingerprint.
- Add tools/threemf_stream.py: parses each part once with lxml iterparse,
resolves each component to the single object it names, and returns each
mesh once plus its placements. The same project now converts in ~8 s at
~380 MB peak (fingerprinting ~190 MB).
- convert_3mf.py and fingerprint.py use it; the GLB stores each object once
and instances repeats. Stats are computed per placement (mirrored
placements keep a positive volume), without concatenating meshes.
- Run sidecar scripts through one queue (lib/python.js) so a preview and
background fingerprinting never load big meshes at the same time;
previews go first, and 3MFs are fingerprinted one per run.
- Surface the converter's JSON error / timeout instead of "Command failed".
- Tests with a Bambu-style component fixture (shared part, mirrored
placement, decoy object) and STL/3MF fingerprint parity.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
