Deploy simready-assets run 20261020-a3cb7f2-captions (GPU variant)

Dockerfile

@@ -1,4 +1,4 @@
-# deploy: 20261020-a3cb7f2 (simready-assets)
+# deploy: 20261020-a3cb7f2-captions (simready-assets)
 # GPU variant of the single-container MVP image for Hugging Face Spaces
 # (Docker SDK): everything docker/Dockerfile.hf-space has (OpenSearch + Neo4j
 # + Redis + deepsearch-api + asset-graph-service + nginx under supervisord),
@@ -185,12 +185,15 @@ RUN python3.12 -m venv /app/.venv \
 
 # The vision checks' prepared asset views on disk (docker/hf-space/check_image_cache.py), on while the
 # CHECK_IMAGE_CACHE_DIR Space Variable is set (pipeline/config.py). The .pth loads it in every venv python. The
-# build fails unless it wraps the vendored llm_client and cached views come back byte-identical (--self-test).
+# build fails unless it wraps the vendored llm_client and cached views come back byte-identical (--self-test), as PNG
+# (the default) and as JPEG (USDSEARCH_VISION_VALIDATION_IMAGE_FORMAT=jpeg, the Space Variable prod runs with).
 COPY docker/hf-space/check_image_cache.py docker/hf-space/check_image_cache.pth /tmp/check-image-hook/
 RUN SP=$(env -u PYTHONPATH /app/.venv/bin/python -c 'import sysconfig; print(sysconfig.get_path("purelib"))') \
     && install -m 644 /tmp/check-image-hook/check_image_cache.py /tmp/check-image-hook/check_image_cache.pth "$SP/" \
     && rm -rf /tmp/check-image-hook \
     && env -u PYTHONPATH CHECK_IMAGE_CACHE_DIR=/tmp/check-image-selftest /app/.venv/bin/python -m check_image_cache --self-test \
+    && env -u PYTHONPATH CHECK_IMAGE_CACHE_DIR=/tmp/check-image-selftest USDSEARCH_VISION_VALIDATION_IMAGE_FORMAT=jpeg \
+        /app/.venv/bin/python -m check_image_cache --self-test \
     && rm -rf /tmp/check-image-selftest
 
 # hf-storage-api — see docker/Dockerfile.hf-space's own copy of this block

docker/hf-space/check_image_cache.py

@@ -1,8 +1,9 @@
 """The vision checks' prepared asset views, kept on disk for every deepsearch-api process of the Space.
 
 Before each check, usdsearch's llm_client prepares the asset's stored view: it decodes the stored thumbnail, fits it
-to USDSEARCH_VISION_VALIDATION_IMAGE_MAX_PX and re-encodes it as base64 PNG (``Validation._encode_reference_images``
--> ``_encode_image``). That is most of a check's API CPU, and a stored view never changes. With
+to USDSEARCH_VISION_VALIDATION_IMAGE_MAX_PX and re-encodes it as base64 PNG, or JPEG with
+USDSEARCH_VISION_VALIDATION_IMAGE_FORMAT=jpeg as on the Space (``Validation._encode_reference_images`` ->
+``_encode_image``). That is most of a check's API CPU, and a stored view never changes. With
 ``CHECK_IMAGE_CACHE_DIR`` set (a Space Variable, pipeline/config.py), each view ``_encode_reference_images`` prepares
 is kept in ``<dir>/<settings>/<sha256 of the stored base64>.b64``: a hit reads the file; a miss prepares the view as
 before and writes it (a temp file, then ``os.replace``), so the API processes share one copy. ``<settings>`` is a
@@ -39,7 +40,7 @@ import zlib
 
 ENV = "CHECK_IMAGE_CACHE_DIR"
 TARGET = "llm_client.validation.validation"
-MAX_BYTES = 2_000_000_000  # ~15k views at ~130 KB (448-px PNG); the Space's ~2,600 stored images take ~340 MB
+MAX_BYTES = 2_000_000_000  # 448-px views: ~28 KB as JPEG, the Space's ~2,600 images ~75 MB (PNG ~121 KB, ~340 MB)
 MIN_FREE = 12_000_000_000  # above the 8 GB the media loader keeps free
 RECHECK_BYTES, RECHECK_S = 64_000_000, 60  # each process re-reads the size and the free space this often
 _DIGEST = re.compile(r"^[0-9a-f]{24}$")
@@ -382,8 +383,9 @@ def sweep(directory, older_than=600):
 
 def self_test():
     """The image build's check (Dockerfile.hf-space-gpu): the .pth loaded this module and it wrapped the vendored
-    llm_client, views read back from the cache are byte-identical to llm_client's own, and the warm-up's imports
-    exist. Raises on any difference. Writes only in CHECK_IMAGE_CACHE_DIR, which the build then removes."""
+    llm_client, views read back from the cache are byte-identical to llm_client's own and in the format the
+    environment names (USDSEARCH_VISION_VALIDATION_IMAGE_FORMAT, unset: PNG), and the warm-up's imports exist.
+    Raises on any difference. Writes only in CHECK_IMAGE_CACHE_DIR, which the build then removes."""
     import base64
     import io
 
@@ -409,6 +411,10 @@ def self_test():
     images = [b64(rgb, "JPEG", quality=90), b64(rgba, "PNG"), b64(rgb.resize((300, 200)), "JPEG", quality=90)]
     config = ValidationSettings(image_max_px=448).to_validation_config()
     plain = [Validation(config=config)._encode_image(image) for image in images]
+    # Wheels that dropped the format setting would pass the build's JPEG run with PNG views.
+    wanted = os.environ.get("USDSEARCH_VISION_VALIDATION_IMAGE_FORMAT") or "png"
+    magic = {"png": "iVBORw0KGgo", "jpeg": "/9j/"}[wanted]
+    assert all(view.startswith(magic) for view in plain), f"the views are not {wanted}"
     miss = Validation(config=config)._encode_reference_images(images)
     reader = Validation(config=config)
     hit = reader._encode_reference_images(images)

docker/hf-space/search_fields.yaml

@@ -15,18 +15,26 @@
 # parser-only (no `label:`) — neither is in nginx.conf.template's
 # railFilters allow-list, so they're invisible to the rail regardless of
 # being registered here; only the LLM query parser picks them up. Verified
-# live via /search/stats/usd_properties: mass is numeric (kg), continuous
-# (1407 distinct values across 1484 of 1488 assets, range 0.001-217652.74,
-# median ~1.77); rigid_body_count is numeric (integer), 37 distinct values
-# across 1484 of 1488 assets, range 1-106, most common value 1 (904
-# assets).
+# live via /search/stats/usd_properties: mass is numeric (kg), continuous;
+# rigid_body_count is numeric (integer), most common value 1.
+#
+# Values and ranges in the descriptions below are the index's (run
+# 20261020-a3cb7f2-captions, 3,599 assets): license 'CC-BY 4.0' 2,513,
+# 'CC-BY-4.0' 1,049, 'cc-by-4.0' 22, 'BSD-3-Clause' 4, 'Apache-2.0' 2,
+# 'LicenseRef-UR-Graphical-Documentation AND BSD-3-Clause' 1, none 8;
+# profiles Robotics-Prop 3,559, Robot-Body 18, AIF-Entity 15, Robot-Gripper 7;
+# mass 0.00013-501,226 kg, median 0.68; rigid bodies 0-126. A property_match
+# eq is an exact, case-sensitive match, so the license (three spellings of
+# CC-BY) is matched with contains (=~*stem*, case-insensitive). They go stale
+# with every re-index: re-check them (docs/development.md, "search_fields.yaml
+# is hand-curated").
 #
 # Sizes (2026-10-14): `dimension` (Central's catalog field, every side) and
 # min_height / max_height / max_size_x / max_size_y write the bounding-box
 # params min_bbox_* / max_bbox_* in metres, so the Explorer shows them as its
 # Dimensions (m) X/Y/Z chips and sliders (the rail's "dimension" control,
-# railFilters in nginx.conf.template). Every asset of the 1,401 carries
-# usd_dimensions, Z up (scaled sides: median 0.25 x 0.29 x 0.18 m, max 12 m).
+# railFilters in nginx.conf.template). Every asset carries usd_dimensions, Z up
+# (2026-10-20, 3,599 assets: scaled sides median 0.19 x 0.17 x 0.11 m, max 33 m).
 # The props face either axis (desk 0.80 x 1.80, counter 0.68 x 3.02, bulk rack
 # 2.45 x 1.23 m), so X and Y are bounded together, once, by the LARGEST
 # horizontal size the query gives (the length, or the width when it is the
@@ -50,10 +58,10 @@ fields:
   operators:
   - eq
   - contains
-  description: SimReady validation profile. The only values are 'Robotics-Prop', 'Robot-Body' and
-    'Robot-Gripper'. Set it ONLY when the query says "profile" or names one of these values. Words for what
+  description: SimReady validation profile. The only values are 'Robotics-Prop', 'Robot-Body', 'Robot-Gripper'
+    and 'AIF-Entity'. Set it ONLY when the query says "profile" or names one of these values. Words for what
     the asset is (robot, gripper, arm, prop, simready) and maker or product names (NVIDIA, LG, Robotiq,
-    Standard Bots, 2F-85, "made by X", "from X") are never a profile; they stay in semantic_query. Another
+    Standard Bots, Vertiv, 2F-85, "made by X", "from X") are never a profile; they stay in semantic_query. Another
     profile name, a profile version ("3.2.0") or a negated profile ("not X", "other than X", "outside X") is
     an unmapped_constraint and no filter.
   examples:
@@ -73,12 +81,16 @@ fields:
   operators:
   - eq
   - contains
-  description: Asset license. The only value is 'CC-BY-4.0' (Creative Commons Attribution 4.0); "CC BY",
-    "CC-BY 4.0" and "creative commons attribution" mean it. Set it ONLY when the query names a license; it is
-    never a stand-in for another constraint (a date, a file type, a format). Another license (MIT, Apache,
-    GPL, proprietary) or a negated license ("not CC-BY") is an unmapped_constraint and no filter.
+  description: Asset license. The index spells one license several ways, so ALWAYS use contains with the
+    short name below, never eq. Creative Commons Attribution 4.0 ("CC BY", "CC-BY 4.0", "creative commons
+    attribution"; almost every asset) -> contains "CC-BY". Apache 2.0 ("Apache", "Apache-2.0") -> contains
+    "Apache". BSD 3-Clause ("BSD", "BSD-3-Clause") -> contains "BSD". Set it ONLY when the query names a
+    license; it is never a stand-in for another constraint (a date, a file type, a format). Any other license
+    (MIT, GPL, proprietary) or a negated license ("not CC-BY") is an unmapped_constraint and no filter.
   examples:
-  - '"CC BY licensed jugs" -> simready_metadata_asset_license eq "CC-BY-4.0", semantic_query "jugs"'
+  - '"CC BY licensed jugs" -> simready_metadata_asset_license contains "CC-BY", semantic_query "jugs"'
+  - '"grippers released under the Apache license" -> simready_metadata_asset_license contains "Apache",
+    semantic_query "grippers"'
   - '"GPL licensed lamps" -> no filter, semantic_query "lamps", unmapped "GPL licensed"'
   - '"glb files of lamps" -> no filter, semantic_query "lamps", unmapped "glb files"'
   label: License
@@ -93,7 +105,7 @@ fields:
   - lt
   - lte
   - eq
-  description: Mass in kilograms (median 1.5 kg, from 0.001 to 217,000). Convert g / 1000, lb x 0.4536, t or
+  description: Mass in kilograms (median 0.7 kg, from 0 to 501,300). Convert g / 1000, lb x 0.4536, t or
     ton x 1000. "heavier than / over N" -> gt N; "under / lighter than N" -> lt N; "no more than / not
     heavier than N" -> lte N; "between A and B" -> gte A and lte B, exactly A and B; only "about / around
     N" -> gte N x 0.9 and lte N x 1.1, never eq. Only a number with a weight unit is a mass; a length, a
@@ -115,8 +127,8 @@ fields:
   - lt
   - lte
   - eq
-  description: Number of rigid bodies (parts that move separately) in the physics setup, an integer from 1 to
-    106; most assets have 1. "single-body" or "one rigid body" -> eq 1; "multi-body" or "several rigid
+  description: Number of rigid bodies (parts that move separately) in the physics setup, an integer from 0 to
+    126; most assets have 1. "single-body" or "one rigid body" -> eq 1; "multi-body" or "several rigid
     bodies" -> gt 1. Only a count of rigid bodies is this field; a size, a mass, a polygon or triangle count
     never is.
   examples:

docker/hf-space/wheels/SOURCE.txt

@@ -1,3 +1,3 @@
 # Generated by scripts/extract_usdsearch_wheels.sh — do not hand-edit.
-source_image=gitlab-master.nvidia.com:5005/omniverse/deeptag/usdsearch/usdsearch:fd61a332
-extracted_at=2026-10-09T14:24:46Z
+source_image=gitlab-master.nvidia.com:5005/omniverse/deeptag/usdsearch/usdsearch:d427fefd
+extracted_at=2026-10-20T20:21:26Z

docker/hf-space/wheels/deepsearch_api-0.1.0-py3-none-any.whl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:380e2f6e9f261b3aedc41c6f3647017002cf68209fed3577eef9e3157209a272
-size 221113
+oid sha256:3ab3d0469632a4e8f31ae2cd19c46f23bb662afabe286266e42520a1aa4c8a32
+size 221753

docker/hf-space/wheels/search_utils-0.1.0-py3-none-any.whl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:35f1cc7cf94f7bf845036cd7332ce2888944eb1c1309532b8863f9e7097c3092
-size 286722
+oid sha256:3e498c1662a82bf7c712e50ed86bb9e6b423dff1f160b113b32b201a2a2b6e4b
+size 287870

docker/hf-space/wheels/siglip2_triton_client-0.6.1-py3-none-any.whl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:daba44a4ac31b836f233e1e7945ab5aaa8b8a2ebcf2702375166f87248f49fc0
+oid sha256:117363f7769e221649ae6e758ba82826ac9e8e00ce227ebfceebf2950d3e8d1e
 size 5482694