unitysvc-data 0.2.8__tar.gz → 0.2.10__tar.gz

{unitysvc_data-0.2.8/src/unitysvc_data.egg-info → unitysvc_data-0.2.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: unitysvc-data
-Version: 0.2.8
+Version: 0.2.10
 Summary: Standard examples and presets for UnitySVC data packages
 Author-email: Bo Peng <bo.peng@unitysvc.com>
 Maintainer-email: Bo Peng <bo.peng@unitysvc.com>

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "unitysvc-data"
-version = "0.2.8"
+version = "0.2.10"
 description = "Standard examples and presets for UnitySVC data packages"
 readme = "README.md"
 authors = [{ name = "Bo Peng", email = "bo.peng@unitysvc.com" }]

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/src/unitysvc_data/_manifest.json

@@ -1992,7 +1992,7 @@
     "llm_code_example_imagetoimage_javascript_v1": {
       "applies_to": {
         "capability": "image-edit",
-        "dialect": "openai"
+        "dialect": "huggingface"
       },
       "category": "code_example",
       "description": "JavaScript example: image-to-image transform via Hugging Face /models/<model>",
@@ -2016,7 +2016,7 @@
     "llm_code_example_imagetoimage_requests_v1": {
       "applies_to": {
         "capability": "image-edit",
-        "dialect": "openai"
+        "dialect": "huggingface"
       },
       "category": "code_example",
       "description": "Python example: image-to-image transform via Hugging Face /models/<model>",
@@ -2043,7 +2043,7 @@
     "llm_code_example_imagetoimage_shell_v1": {
       "applies_to": {
         "capability": "image-edit",
-        "dialect": "openai"
+        "dialect": "huggingface"
       },
       "category": "code_example",
       "description": "Bash example: image-to-image transform via Hugging Face /models/<model> using curl",
@@ -2067,7 +2067,7 @@
     "llm_code_example_imagetoimage_shell_v2": {
       "applies_to": {
         "capability": "image-edit",
-        "dialect": "openai"
+        "dialect": "huggingface"
       },
       "category": "code_example",
       "description": "Bash example: image-to-image transform via Hugging Face /models/<model> using curl",
@@ -3160,7 +3160,7 @@
     "llm_code_example_ttv_javascript_v1": {
       "applies_to": {
         "capability": "video-generate",
-        "dialect": "openai"
+        "dialect": "huggingface"
       },
       "category": "code_example",
       "description": "JavaScript example: text-to-video via HF /models/<model>",
@@ -3181,7 +3181,7 @@
     "llm_code_example_ttv_requests_v1": {
       "applies_to": {
         "capability": "video-generate",
-        "dialect": "openai"
+        "dialect": "huggingface"
       },
       "category": "code_example",
       "description": "Python example: text-to-video via HF /models/<model>",
@@ -3205,7 +3205,7 @@
     "llm_code_example_ttv_shell_v1": {
       "applies_to": {
         "capability": "video-generate",
-        "dialect": "openai"
+        "dialect": "huggingface"
       },
       "category": "code_example",
       "description": "Bash example: text-to-video via HF /models/<model> using curl",
@@ -3226,7 +3226,7 @@
     "llm_code_example_ttv_shell_v2": {
       "applies_to": {
         "capability": "video-generate",
-        "dialect": "openai"
+        "dialect": "huggingface"
       },
       "category": "code_example",
       "description": "Bash example: text-to-video via HF /models/<model> using curl",

unitysvc_data-0.2.10/src/unitysvc_data/classifiers.py

@@ -0,0 +1,317 @@
+"""The classifier registry — the values ``applies_to`` may use, and how each reads.
+
+``applies_to`` decides which examples a service gets (see ``tools/build.py``).
+Its four axes used to be documented only in a comment and validated nowhere, so
+a typo silently *widened* a selector rather than narrowing it: an absent key
+means "no constraint", so ``capabilty = "chat"`` made an example apply to every
+service. One such omission had to be found by hand and fixed in 0.2.3
+(``llm/code-example-tts-shell``, which named a capability but no dialect).
+
+This module is the single declaration of what each axis denotes and which values
+it accepts. ``tools/build.py`` rejects anything else, with a near-miss hint.
+
+It also carries each value's **display label**, which is what ``presets._title``
+renders. Keeping the labels here rather than in private tables next to the title
+builder buys two things:
+
+* a value cannot be used without being registered, so no raw token can reach a
+  customer-facing title (``dashscope_audio_task`` did exactly that: eight titles
+  read ``(speech, dashscope_audio_task)``);
+* an empty label becomes declared data rather than a hardcoded skip, so "why is
+  ``openai`` invisible in titles?" is answered next to the value.
+
+A title is a document's KEY -- the backend upserts on
+``(entity_id, context_type, title)``. Two values that share a label therefore
+merge two documents into one, and the loser is silently overwritten. That is why
+labels must stay distinct wherever the values they name can appear on one
+service, and why ``tests/test_classifiers.py`` asserts it against the real
+selector rather than trusting this docstring.
+"""
+
+from __future__ import annotations
+
+import difflib
+from typing import NamedTuple
+
+
+class Classifier(NamedTuple):
+    """One allowed value of one axis."""
+
+    #: How the value reads inside a title's parenthesised qualifier list.
+    #: Empty means "contributes nothing" -- a deliberate choice, explained in
+    #: ``note``, not an oversight.
+    label: str
+    #: Rendered as ``"<label> input"``: the label names a request *dialect* the
+    #: caller writes, not a client library. ``cohere`` is a named SDK and reads
+    #: as itself; ``anthropic`` is a wire shape and reads as "Anthropic-style
+    #: input".
+    caller_dialect: bool = False
+    #: Why the label is empty, or anything else a reader needs. Documentation
+    #: only; nothing reads it.
+    note: str = ""
+
+
+# --------------------------------------------------------------------------- #
+# capability -- what the example demonstrates. The offering declares a LIST of
+# these, and a collection fans out over every one, so two presets with
+# different capabilities both land in one ``documents`` mapping. Their titles
+# must differ: an omni model declaring chat, text-to-speech and speech-to-text
+# used to ship ONE of those three sets, the other two overwritten by title.
+# --------------------------------------------------------------------------- #
+CAPABILITIES: dict[str, Classifier] = {
+    "chat": Classifier("", note="A bare title already reads as chat; labelling it would churn nearly every published service for no gain."),
+    "image-text-to-text": Classifier("", note="Carried by the `vision` feature bit instead, so it cannot collide with a labelled sibling."),
+    "embed": Classifier("embeddings"),
+    "rerank": Classifier("rerank"),
+    "moderate": Classifier("moderation"),
+    "image-generate": Classifier("image"),
+    "image-edit": Classifier("image edit"),
+    "video-generate": Classifier("video"),
+    "speech-to-text": Classifier("transcription"),
+    "text-to-speech": Classifier("speech"),
+}
+
+# --------------------------------------------------------------------------- #
+# dialect -- the request shape the CALLER writes. A service's `input_formats` is
+# a SET, so two presets with different dialects can both be selected for one
+# service; their labels must therefore stay distinct.
+#
+# The DashScope tokens are finer than the `dashscope` request format, because
+# they name different native endpoints and body shapes: text-generation,
+# multimodal-generation, and the audio-task surface. A model serves one of them,
+# so a service does not normally declare two -- but `text` and `multimodal`
+# shared the label "DashScope" and so rendered identical titles, which is a
+# silent overwrite waiting for the first service that does.
+# --------------------------------------------------------------------------- #
+DIALECTS: dict[str, Classifier] = {
+    "openai": Classifier("", note="The platform's default dialect: naming it in every title would say nothing."),
+    "anthropic": Classifier("Anthropic-style", caller_dialect=True),
+    "cohere": Classifier("Cohere SDK"),
+    "cerebras": Classifier("Cerebras SDK"),
+    "bedrock_converse": Classifier("boto3 Converse"),
+    "bedrock_invoke": Classifier("boto3 InvokeModel"),
+    # Named for the surface, not for one model family: the same
+    # ``/models/<model>`` inference API serves sentence embeddings, image
+    # editing and text-to-video. It was "sentence-transformers", which could
+    # not honestly title the image-edit and video examples that also speak it.
+    "huggingface": Classifier("Hugging Face"),
+    "dashscope": Classifier("DashScope", caller_dialect=True),
+    "dashscope_text": Classifier("DashScope text", caller_dialect=True),
+    "dashscope_multimodal": Classifier("DashScope", caller_dialect=True),
+    "dashscope_audio_task": Classifier("DashScope audio task"),
+}
+
+# --------------------------------------------------------------------------- #
+# upstream -- the dialect the service's upstream speaks, which differs from
+# `dialect` when the gateway translates.
+#
+# Every label here is empty, and unlike the other axes that is not a style
+# choice: a collection is built with exactly ONE upstream
+# (`source["upstream_dialect"]`), so two presets that declare different
+# upstreams can never be selected together and cannot collide however they are
+# titled. Giving the axis a label would add "via OpenAI" to ~80% of titles to
+# separate documents that are already separate.
+# --------------------------------------------------------------------------- #
+UPSTREAMS: dict[str, Classifier] = {
+    "openai": Classifier("", note="One upstream per collection, so this axis cannot cause a title collision."),
+    "anthropic": Classifier("", note="See `openai`."),
+    "dashscope": Classifier("", note="See `openai`."),
+}
+
+# --------------------------------------------------------------------------- #
+# feature -- an attribute the service must advertise before the example applies.
+# Passed as a SET, so these can co-occur and must stay distinct.
+# --------------------------------------------------------------------------- #
+FEATURES: dict[str, Classifier] = {
+    "streaming": Classifier("streaming"),
+    "tools": Classifier("tools"),
+    "vision": Classifier("vision"),
+}
+
+class Axis(NamedTuple):
+    """One ``applies_to`` key: its allowed values, and how it behaves."""
+
+    values: dict[str, Classifier]
+    #: Whether two examples differing ONLY on this axis can both end up in one
+    #: service's ``documents`` mapping. When they can, their titles MUST differ
+    #: or one silently overwrites the other -- which makes this flag the input
+    #: to the collision check, not documentation.
+    co_occurs: bool
+    note: str = ""
+
+
+#: Axis name -> the axis. The keys are exactly the keys ``applies_to`` accepts.
+AXES: dict[str, Axis] = {
+    "capability": Axis(
+        CAPABILITIES,
+        co_occurs=True,
+        note="A collection fans out over every capability the offering declares, "
+        "and each iteration adds to the SAME documents mapping, so two "
+        "capabilities meet there.",
+    ),
+    "dialect": Axis(
+        DIALECTS,
+        co_occurs=True,
+        note="`input_formats` is a set, so one service can be served examples "
+        "for several dialects at once.",
+    ),
+    "upstream": Axis(
+        UPSTREAMS,
+        co_occurs=False,
+        note="A collection is built with exactly ONE upstream "
+        "(`source['upstream_dialect']`), so two examples declaring different "
+        "upstreams are never selected together and cannot collide however "
+        "they are titled. This is why every upstream label is empty.",
+    ),
+    "feature": Axis(
+        FEATURES,
+        co_occurs=True,
+        note="Matched against a set, so a streaming example and a tools "
+        "example can both apply.",
+    ),
+}
+
+#: ``(axis, value)`` pairs whose empty label is deliberate, on an axis whose
+#: values otherwise MUST be labelled. Each needs a reachability argument,
+#: because two unlabelled values on a co-occurring axis render one title:
+#:
+#: * ``chat`` and ``image-text-to-text`` are both unlabelled, and would collide
+#:   -- except that declaring ``image-text-to-text`` implies the ``vision``
+#:   feature (see ``presets._VISION_CAPABILITIES``), so its examples always
+#:   carry the ``vision`` qualifier and the pair is unreachable.
+#: * ``openai`` is the platform's default dialect; naming it in every title
+#:   would say nothing.
+#:
+#: The corpus check is what proves these arguments still hold. Adding a pair
+#: here without one is how a silent overwrite gets introduced.
+UNLABELLED: frozenset[tuple[str, str]] = frozenset(
+    {
+        ("capability", "chat"),
+        ("capability", "image-text-to-text"),
+        ("dialect", "openai"),
+    }
+)
+
+
+#: Label texts an axis may deliberately reuse, with the reachability argument
+#: that makes the reuse safe. Same contract as ``UNLABELLED``: a sharing that
+#: is not declared here fails the registry check.
+#:
+#: * ``dashscope`` and ``dashscope_multimodal`` both read "DashScope" because
+#:   they name the same provider, and no capability has examples on both of
+#:   them (``dashscope`` is embeddings-only; ``dashscope_multimodal`` carries
+#:   chat, transcription and speech). Their titles therefore always differ by
+#:   the capability qualifier. Adding an embeddings example for
+#:   ``dashscope_multimodal`` would break that, and the corpus check -- not
+#:   this list -- is what would catch it.
+SHARED_LABELS: frozenset[tuple[str, str]] = frozenset(
+    {
+        ("dialect", "DashScope"),
+    }
+)
+
+
+def co_occurs(axis: str) -> bool:
+    """Whether two examples differing only on this axis can meet on one service.
+
+    Unknown axes are treated as co-occurring: the safe default is to require
+    distinct titles.
+    """
+    entry = AXES.get(axis)
+    return entry.co_occurs if entry else True
+
+
+def check_registry() -> list[str]:
+    """Problems with the registry itself, independent of any example.
+
+    Stated over the declared values rather than the ones in use, so a value
+    added with a duplicate or missing label fails before any example adopts it.
+    """
+    problems: list[str] = []
+    for axis_name, axis in sorted(AXES.items()):
+        if not axis.co_occurs:
+            continue  # titles may coincide; nothing to keep distinct
+        by_label: dict[str, list[str]] = {}
+        for value, entry in sorted(axis.values.items()):
+            if not entry.label:
+                if (axis_name, value) not in UNLABELLED:
+                    problems.append(
+                        f"{axis_name} {value!r} has no display label, so two examples "
+                        f"differing only in {axis_name} would render one title and one "
+                        f"would silently overwrite the other. Give it a label, or add it "
+                        f"to UNLABELLED with the reason the clash is unreachable."
+                    )
+                continue
+            by_label.setdefault(entry.label, []).append(value)
+        for label_text, values in sorted(by_label.items()):
+            if len(values) > 1 and (axis_name, label_text) not in SHARED_LABELS:
+                problems.append(
+                    f"{axis_name} values {values} share the display label "
+                    f"{label_text!r}; their titles would be identical. A title is a "
+                    f"document's key, so the loser is overwritten at ingest. If the "
+                    f"clash is unreachable, add it to SHARED_LABELS with the reason."
+                )
+    return problems
+
+
+def classifier(axis: str, value: str) -> Classifier:
+    """The registry entry for one value.
+
+    Raises:
+        KeyError: The axis or the value is not registered. Callers that reach
+            here have skipped :func:`validate`, which is the layer meant to
+            produce a readable message.
+    """
+    return AXES[axis].values[value]
+
+
+def label(axis: str, value: str | None) -> str:
+    """How a value reads in a title, or ``""`` when it contributes nothing.
+
+    An unregistered value returns ``""`` rather than raising, so a title is
+    never built out of a raw token. Validation is what rejects it; this only
+    refuses to leak it.
+    """
+    if value is None:
+        return ""
+    axis_entry = AXES.get(axis)
+    entry = axis_entry.values.get(value) if axis_entry else None
+    return entry.label if entry else ""
+
+
+def is_caller_dialect(value: str | None) -> bool:
+    """Whether a dialect's label names a wire shape rather than a client."""
+    entry = DIALECTS.get(value or "")
+    return bool(entry and entry.caller_dialect)
+
+
+def validate(applies_to: dict[str, object]) -> list[str]:
+    """Every problem with one ``applies_to`` block, as readable messages.
+
+    Returns an empty list when the block is valid. Unknown keys are reported
+    as well as unknown values, because an absent key means "no constraint": a
+    misspelled axis does not narrow the selector, it silently widens it.
+    """
+    problems: list[str] = []
+    for axis, value in sorted(applies_to.items()):
+        axis_entry = AXES.get(axis)
+        if axis_entry is None:
+            hint = difflib.get_close_matches(axis, AXES, n=1)
+            suffix = f" Did you mean {hint[0]!r}?" if hint else ""
+            problems.append(
+                f"unknown applies_to key {axis!r} -- an unrecognised key is not a "
+                f"constraint, so the example would apply to every service.{suffix}"
+            )
+            continue
+        if not isinstance(value, str):
+            problems.append(f"applies_to.{axis} must be a string, got {type(value).__name__}")
+            continue
+        if value not in axis_entry.values:
+            hint = difflib.get_close_matches(value, axis_entry.values, n=1)
+            suffix = f" Did you mean {hint[0]!r}?" if hint else ""
+            problems.append(
+                f"unknown {axis} {value!r}. Register it in "
+                f"src/unitysvc_data/classifiers.py (with the label it should "
+                f"read as in a title) or fix the spelling.{suffix}"
+            )
+    return problems

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/src/unitysvc_data/examples/llm/code-example-imagetoimage-javascript/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Image to image", min_expected_metrics = { images_generated = 1 } }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "image-edit", dialect = "openai" }
+applies_to = { capability = "image-edit", dialect = "huggingface" }
 +++
 
 # llm / code-example-imagetoimage-javascript — image-to-image via `fetch`
@@ -33,6 +33,12 @@ Optional:
 ## Versions
 
 ### v1 — initial release
+- **Amended in 0.2.10**: `dialect` corrected to `huggingface`. The 0.2.3
+  amendment below declared `openai`, but this template posts to the Hugging
+  Face inference API (`/models/<model>`) -- so the example was offered to
+  every OpenAI-dialect service, whose endpoint does not serve that path.
+  Selection metadata only; the template is unchanged. The document title
+  gains a `Hugging Face` qualifier as a result.
 - **Amended in 0.2.3**: `applies_to` now declares `dialect = "openai"`.
   It named only the capability, and an absent key means "no constraint" — so
   this example applied to callers writing any dialect, including ones whose

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/src/unitysvc_data/examples/llm/code-example-imagetoimage-requests/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Image to image", requirements = ["requests"], min_expected_metrics = { images_generated = 1 } }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "image-edit", dialect = "openai" }
+applies_to = { capability = "image-edit", dialect = "huggingface" }
 +++
 
 # llm / code-example-imagetoimage-requests — image-to-image via `requests`
@@ -34,6 +34,12 @@ Optional:
 ## Versions
 
 ### v1 — initial release
+- **Amended in 0.2.10**: `dialect` corrected to `huggingface`. The 0.2.3
+  amendment below declared `openai`, but this template posts to the Hugging
+  Face inference API (`/models/<model>`) -- so the example was offered to
+  every OpenAI-dialect service, whose endpoint does not serve that path.
+  Selection metadata only; the template is unchanged. The document title
+  gains a `Hugging Face` qualifier as a result.
 - **Amended in 0.2.3**: `applies_to` now declares `dialect = "openai"`.
   It named only the capability, and an absent key means "no constraint" — so
   this example applied to callers writing any dialect, including ones whose

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/src/unitysvc_data/examples/llm/code-example-imagetoimage-shell/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Image to image", min_expected_metrics = { images_generated = 1 } }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "image-edit", dialect = "openai" }
+applies_to = { capability = "image-edit", dialect = "huggingface" }
 +++
 
 # llm / code-example-imagetoimage-shell — image-to-image via `curl`
@@ -44,6 +44,12 @@ Optional:
   stays out of the published example; the unconditional success marker is gone.
 - **Amended in 0.2.2.** Earlier installs of this package ship a v2 whose
   verification lived in metadata instead.
+- **Amended in 0.2.10**: `dialect` corrected to `huggingface`. The 0.2.3
+  amendment below declared `openai`, but this template posts to the Hugging
+  Face inference API (`/models/<model>`) -- so the example was offered to
+  every OpenAI-dialect service, whose endpoint does not serve that path.
+  Selection metadata only; the template is unchanged. The document title
+  gains a `Hugging Face` qualifier as a result.
 - **Amended in 0.2.3**: `applies_to` now declares `dialect = "openai"`.
   It named only the capability, and an absent key means "no constraint" — so
   this example applied to callers writing any dialect, including ones whose

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/src/unitysvc_data/examples/llm/code-example-ttv-javascript/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Text to video" }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "video-generate", dialect = "openai" }
+applies_to = { capability = "video-generate", dialect = "huggingface" }
 +++
 
 # llm / code-example-ttv-javascript — text-to-video via `fetch`
@@ -33,6 +33,12 @@ Optional:
 ## Versions
 
 ### v1 — initial release
+- **Amended in 0.2.10**: `dialect` corrected to `huggingface`. The 0.2.3
+  amendment below declared `openai`, but this template posts to the Hugging
+  Face inference API (`/models/<model>`) -- so the example was offered to
+  every OpenAI-dialect service, whose endpoint does not serve that path.
+  Selection metadata only; the template is unchanged. The document title
+  gains a `Hugging Face` qualifier as a result.
 - **Amended in 0.2.3**: `applies_to` now declares `dialect = "openai"`.
   It named only the capability, and an absent key means "no constraint" — so
   this example applied to callers writing any dialect, including ones whose

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/src/unitysvc_data/examples/llm/code-example-ttv-requests/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Text to video", requirements = ["requests"] }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "video-generate", dialect = "openai" }
+applies_to = { capability = "video-generate", dialect = "huggingface" }
 +++
 
 # llm / code-example-ttv-requests — text-to-video via `requests`
@@ -33,6 +33,12 @@ Optional:
 ## Versions
 
 ### v1 — initial release
+- **Amended in 0.2.10**: `dialect` corrected to `huggingface`. The 0.2.3
+  amendment below declared `openai`, but this template posts to the Hugging
+  Face inference API (`/models/<model>`) -- so the example was offered to
+  every OpenAI-dialect service, whose endpoint does not serve that path.
+  Selection metadata only; the template is unchanged. The document title
+  gains a `Hugging Face` qualifier as a result.
 - **Amended in 0.2.3**: `applies_to` now declares `dialect = "openai"`.
   It named only the capability, and an absent key means "no constraint" — so
   this example applied to callers writing any dialect, including ones whose

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/src/unitysvc_data/examples/llm/code-example-ttv-shell/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Text to video" }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "video-generate", dialect = "openai" }
+applies_to = { capability = "video-generate", dialect = "huggingface" }
 +++
 
 # llm / code-example-ttv-shell — text-to-video via `curl`
@@ -44,6 +44,12 @@ Optional:
   stays out of the published example; the unconditional success marker is gone.
 - **Amended in 0.2.2.** Earlier installs of this package ship a v2 whose
   verification lived in metadata instead.
+- **Amended in 0.2.10**: `dialect` corrected to `huggingface`. The 0.2.3
+  amendment below declared `openai`, but this template posts to the Hugging
+  Face inference API (`/models/<model>`) -- so the example was offered to
+  every OpenAI-dialect service, whose endpoint does not serve that path.
+  Selection metadata only; the template is unchanged. The document title
+  gains a `Hugging Face` qualifier as a result.
 - **Amended in 0.2.3**: `applies_to` now declares `dialect = "openai"`.
   It named only the capability, and an absent key means "no constraint" — so
   this example applied to callers writing any dialect, including ones whose

{unitysvc_data-0.2.8 → unitysvc_data-0.2.10}/src/unitysvc_data/examples/llm/dashscope-omni-tts-connectivity/connectivity-v1.sh.j2

@@ -2,6 +2,13 @@
 # DashScope connectivity probe (omni text-to-speech): ask for audio output and
 # assert the stream really carries an audio frame. A 2xx alone is not enough --
 # without the SSE header the call still succeeds and returns no audio at all.
+#
+# Nothing here pipes "$response" into another command. The stream is ~140 KB of
+# base64 audio, far past the pipe buffer, so a reader that stops early --
+# `head -c 200`, or `grep -q` on its first match -- closes the pipe and the
+# writer takes SIGPIPE. Under `pipefail` that surfaces as exit 141 and fails a
+# probe whose request succeeded. Bash substring and `case` read the variable
+# directly instead.
 set -e -o pipefail
 
 : "${UNITYSVC_API_KEY:?UNITYSVC_API_KEY is not set}"
@@ -12,10 +19,10 @@ response=$(curl --fail-with-body -sS -N "{{ service_base_url }}" \
   -H "X-DashScope-SSE: enable" \
   -d '{"model": "{{ routing_key.model }}", "input": {"messages": [{"role": "user", "content": [{"text": "Say this is a test"}]}]}, "parameters": {"modalities": ["text", "audio"], "incremental_output": true}}')
 
-printf '%s' "$response" | head -c 200
+printf '%s\n' "${response:0:200}"
 {%- if not customer_display %}
-if ! printf '%s' "$response" | grep -q '"audio"'; then
-  echo 'unexpected response: no "audio" in the stream' >&2
-  exit 1
-fi
+case "$response" in
+  *'"audio"'*) ;;
+  *) echo 'unexpected response: no "audio" in the stream' >&2; exit 1 ;;
+esac
 {%- endif %}