Spaces:
Sleeping
Sleeping
Commit History
-
spacing button fixes
9741d0f
-
second pass at SDK update
c2e4ea5 -
spaces update
f7b1118 -
stylistic fixes
d4638ee
-
Fix example-query clearing, force query_string, lazy-load CLIP
Route every example / "See more like this" click through a shared _reset_query() that clears all signal toggles, query boxes, filter rows, and the stored similar-vector before applying the new query — so stale signals no longer fuse into the next search. It also forces fts_type=query_string, so the Lucene-syntax FTS examples always run under query_string even if the user had switched the dropdown to text. Also gate _load_clip() behind the dense-text path: it was called for every search, pulling in torch/CLIP even for FTS-only, sparse-only, and similar-vector queries that never use it (and hard-crashing in a torch-less env). CLIP now loads only when embedding a dense text query. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>0d9303a
-
Slim Streamlit deps for HF Spaces build
Pin CPU-only torch/torchvision (default wheels pull ~2GB of CUDA libs), drop deps the running app never imports (jupyter/ipython/matplotlib and the deferred-sound stack transformers/soundfile/soxr). Also fix the DEPLOY.md upload command to drop --resize (panels are already small; resize inflates). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>4191ae4 -
Distill streamlit-hosting: HF Spaces config + deploy guide
Add HF Spaces front-matter to README, rewrite DEPLOY.md for the Streamlit-on-HF-Spaces path, and drop the Vercel-only architecture diagram. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>c3536b3 -
Vercel architecture docs (DEPLOY.md + excalidraw diagram)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>1542156 -
S3 image hosting + unwire sound button (shared hosting foundation)
Add src/storage/images.py to resolve panel image_path to an S3/CDN URL (IMAGE_STORE=s3) or local base64 (default), and scripts/upload_images_to_s3.py to push panels to S3 with optional downscaling and immutable cache headers. app.py now renders via image_src() and the 'Find sounds' button is unwired (backend kept). Adds boto3 and .env.example. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>a3d7de8
-
updated git ignore
ac89c28
-
added personable blog
11b05af
-
added sounds
04a743f -
blog updates
fb1b99d
-
audio pipeline
d0ba735 -
readme changes
73e77a4 -
add blog
eeff165 -
before sound
0f75049 -
streamlit app
830fbdc
-
mps/cuda and delete example
5084ad7 -
restructure project: pivot from multi-approach notebooks to focused COMICS panel search build
The original scaffold explored three separate retrieval approaches across six numbered notebooks (text+dense, image embeddings, hybrid) with a flat utils/ package. This commit replaces that structure with a production-oriented layout targeting the COMICS dataset specifically. Key changes: - Removed numbered exploration notebooks (00–06) and flat utils/ package - Added src/ package with modules for ingest, embeddings, pinecone, and search - ingest/: manifest building and OCR text cleaning - embeddings/: dense image embeddings (embed_images) and sparse text (embed_sparse_text) - pinecone/: index creation, document assembly, and panel upserting - search/: dense, sparse, FTS, and combined search with client-side fusion - Added notebooks/query_comics_index.ipynb as the single interactive entry point - Added config.yaml for shared runtime configuration - Added updated-build.md capturing the revised build spec and data source decisions (COMICS dataset via umiacs.umd.edu, Pinecone d…3bf4df9 -
add .gitignore for Python/Jupyter project
Ignores .env secrets, .venv/, __pycache__, .ipynb_checkpoints, data/, and macOS artifacts. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c10184e
-
scaffold all notebooks, utils, and project structure
45f5822 -
update plan with finalized index schema and embedding decisions
9c0e468 -
add project plan
e6419b4 -
initial commit
7a49b4f