unitysvc-data 0.2.11__tar.gz → 0.2.13__tar.gz

{unitysvc_data-0.2.11/src/unitysvc_data.egg-info → unitysvc_data-0.2.13}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: unitysvc-data
-Version: 0.2.11
+Version: 0.2.13
 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-0.2.11 → unitysvc_data-0.2.13}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "unitysvc-data"
-version = "0.2.11"
+version = "0.2.13"
 description = "Standard examples and presets for UnitySVC data packages"
 readme = "README.md"
 authors = [{ name = "Bo Peng", email = "bo.peng@unitysvc.com" }]

{unitysvc_data-0.2.11 → unitysvc_data-0.2.13}/src/unitysvc_data/_manifest.json

@@ -99,7 +99,15 @@
     "llm_connectivity_tts": "llm_connectivity_tts_v1",
     "llm_connectivity_tts_dashscope": "llm_connectivity_tts_dashscope_v1",
     "llm_description": "llm_description_v2",
-    "llm_request_template": "llm_request_template_v2",
+    "llm_request_template": "llm_request_template_v4",
+    "llm_request_template_embed": "llm_request_template_embed_v1",
+    "llm_request_template_image_generate": "llm_request_template_image_generate_v1",
+    "llm_request_template_image_text_to_text": "llm_request_template_image_text_to_text_v1",
+    "llm_request_template_moderate": "llm_request_template_moderate_v1",
+    "llm_request_template_rerank": "llm_request_template_rerank_v1",
+    "llm_request_template_speech_to_text": "llm_request_template_speech_to_text_v1",
+    "llm_request_template_text_to_speech": "llm_request_template_text_to_speech_v1",
+    "llm_request_template_video_generate": "llm_request_template_video_generate_v1",
     "msg_request_template": "msg_request_template_v1",
     "msg_to_apprise_code_example_py": "msg_to_apprise_code_example_py_v1",
     "msg_to_apprise_code_example_sh": "msg_to_apprise_code_example_sh_v1",
@@ -3764,6 +3772,130 @@
       "source_readme": "llm/description/README.md",
       "version": 2
     },
+    "llm_request_template_embed_v1": {
+      "applies_to": {
+        "capability": "embed"
+      },
+      "category": "request_template",
+      "description": "Minimal embeddings request body for each request format",
+      "example_file": "llm/request-template-embed/request-template-embed-v1.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_request_template_embed",
+      "source_readme": "llm/request-template-embed/README.md",
+      "version": 1
+    },
+    "llm_request_template_image_generate_v1": {
+      "applies_to": {
+        "capability": "image-generate"
+      },
+      "category": "request_template",
+      "description": "Minimal image generation request body for each request format",
+      "example_file": "llm/request-template-image-generate/request-template-image-generate-v1.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_request_template_image_generate",
+      "source_readme": "llm/request-template-image-generate/README.md",
+      "version": 1
+    },
+    "llm_request_template_image_text_to_text_v1": {
+      "applies_to": {
+        "capability": "image-text-to-text"
+      },
+      "category": "request_template",
+      "description": "Minimal vision (image and text) request body for each request format",
+      "example_file": "llm/request-template-image-text-to-text/request-template-image-text-to-text-v1.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_request_template_image_text_to_text",
+      "source_readme": "llm/request-template-image-text-to-text/README.md",
+      "version": 1
+    },
+    "llm_request_template_moderate_v1": {
+      "applies_to": {
+        "capability": "moderate"
+      },
+      "category": "request_template",
+      "description": "Minimal moderation request body for each request format",
+      "example_file": "llm/request-template-moderate/request-template-moderate-v1.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_request_template_moderate",
+      "source_readme": "llm/request-template-moderate/README.md",
+      "version": 1
+    },
+    "llm_request_template_rerank_v1": {
+      "applies_to": {
+        "capability": "rerank"
+      },
+      "category": "request_template",
+      "description": "Minimal rerank request body for each request format",
+      "example_file": "llm/request-template-rerank/request-template-rerank-v1.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_request_template_rerank",
+      "source_readme": "llm/request-template-rerank/README.md",
+      "version": 1
+    },
+    "llm_request_template_speech_to_text_v1": {
+      "applies_to": {
+        "capability": "speech-to-text"
+      },
+      "category": "request_template",
+      "description": "Minimal transcription request body for each request format",
+      "example_file": "llm/request-template-speech-to-text/request-template-speech-to-text-v1.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {},
+      "preset_name": "llm_request_template_speech_to_text",
+      "source_readme": "llm/request-template-speech-to-text/README.md",
+      "version": 1
+    },
+    "llm_request_template_text_to_speech_v1": {
+      "applies_to": {
+        "capability": "text-to-speech"
+      },
+      "category": "request_template",
+      "description": "Minimal speech synthesis request body for each request format",
+      "example_file": "llm/request-template-text-to-speech/request-template-text-to-speech-v1.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_request_template_text_to_speech",
+      "source_readme": "llm/request-template-text-to-speech/README.md",
+      "version": 1
+    },
     "llm_request_template_v1": {
       "applies_to": {
         "capability": "chat"
@@ -3775,7 +3907,9 @@
       "is_public": true,
       "meta": {},
       "mime_type": "json",
-      "parameters": {},
+      "parameters": {
+        "version_prefix": "/v1"
+      },
       "preset_name": "llm_request_template",
       "source_readme": "llm/request-template/README.md",
       "version": 1
@@ -3791,11 +3925,65 @@
       "is_public": true,
       "meta": {},
       "mime_type": "json",
-      "parameters": {},
+      "parameters": {
+        "version_prefix": "/v1"
+      },
       "preset_name": "llm_request_template",
       "source_readme": "llm/request-template/README.md",
       "version": 2
     },
+    "llm_request_template_v3": {
+      "applies_to": {
+        "capability": "chat"
+      },
+      "category": "request_template",
+      "description": "Minimal chat request body for each request format",
+      "example_file": "llm/request-template/request-template-v3.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_request_template",
+      "source_readme": "llm/request-template/README.md",
+      "version": 3
+    },
+    "llm_request_template_v4": {
+      "applies_to": {
+        "capability": "chat"
+      },
+      "category": "request_template",
+      "description": "Minimal chat request body for each request format",
+      "example_file": "llm/request-template/request-template-v4.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_request_template",
+      "source_readme": "llm/request-template/README.md",
+      "version": 4
+    },
+    "llm_request_template_video_generate_v1": {
+      "applies_to": {
+        "capability": "video-generate"
+      },
+      "category": "request_template",
+      "description": "Minimal video generation request body for each request format",
+      "example_file": "llm/request-template-video-generate/request-template-video-generate-v1.json",
+      "is_active": true,
+      "is_public": true,
+      "meta": {},
+      "mime_type": "json",
+      "parameters": {},
+      "preset_name": "llm_request_template_video_generate",
+      "source_readme": "llm/request-template-video-generate/README.md",
+      "version": 1
+    },
     "msg_request_template_v1": {
       "applies_to": {},
       "category": "request_template",

{unitysvc_data-0.2.11 → unitysvc_data-0.2.13}/src/unitysvc_data/classifiers.py

@@ -49,6 +49,13 @@ class Classifier(NamedTuple):
     #: Why the label is empty, or anything else a reader needs. Documentation
     #: only; nothing reads it.
     note: str = ""
+    #: The ``feature`` that names this value when its own ``label`` is empty.
+    #: ``titles.title`` renders that feature's label for a document of this value
+    #: whether or not the document declares the feature itself, so the title is
+    #: told apart from a bare sibling by construction rather than by every
+    #: document remembering to say so. Only meaningful on an unlabelled value;
+    #: ``check_registry`` rejects a carrier that renders nothing.
+    carried_by: str = ""
 
 
 # --------------------------------------------------------------------------- #
@@ -60,7 +67,13 @@ class Classifier(NamedTuple):
 # --------------------------------------------------------------------------- #
 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."),
+    "image-text-to-text": Classifier(
+        "",
+        carried_by="vision",
+        note="Carried by the `vision` feature bit instead, so it cannot collide with a labelled sibling. "
+        "The title builder applies that bit itself (see `carried_by`), because a document with no "
+        "feature of its own -- a request template -- would otherwise render exactly like chat.",
+    ),
     "embed": Classifier("embeddings"),
     "rerank": Classifier("rerank"),
     "moderate": Classifier("moderation"),
@@ -176,9 +189,16 @@ AXES: dict[str, Axis] = {
 #: 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.
+#:   -- except that ``image-text-to-text`` is *carried by* the ``vision``
+#:   feature: ``titles.title`` renders that qualifier for every document of the
+#:   capability, whatever category it is and whether or not it declares the
+#:   feature, so the pair is distinct by construction. (This used to rest on
+#:   every image-text-to-text EXAMPLE declaring ``feature = "vision"``, which a
+#:   request template -- declaring no feature at all -- does not. The reasoning
+#:   held for examples and failed for the first document that was not one.)
+#:   What construction does not cover is a CHAT document that itself declares
+#:   ``feature = "vision"``: it would render like an image-text-to-text one, and
+#:   the corpus check in ``tools/build.py`` is what rejects that.
 #: * ``openai`` is the platform's default dialect; naming it in every title
 #:   would say nothing.
 #:
@@ -229,8 +249,41 @@ def check_registry() -> list[str]:
     """
     problems: list[str] = []
     for axis_name, axis in sorted(AXES.items()):
+        for value, entry in sorted(axis.values.items()):
+            if not entry.carried_by:
+                continue
+            carrier = FEATURES.get(entry.carried_by)
+            if entry.label:
+                problems.append(
+                    f"{axis_name} {value!r} has both the label {entry.label!r} and "
+                    f"carried_by {entry.carried_by!r}. carried_by stands in for a label "
+                    f"the value does not have; with one, it only adds a second qualifier."
+                )
+            if carrier is None or not carrier.label:
+                problems.append(
+                    f"{axis_name} {value!r} is carried_by {entry.carried_by!r}, which "
+                    f"{'is not a registered feature' if carrier is None else 'has no label'}, "
+                    f"so it would render nothing and the value would still be bare."
+                )
         if not axis.co_occurs:
             continue  # titles may coincide; nothing to keep distinct
+        # A value is BARE when nothing names it: no label, and no feature carrying
+        # it. Two bare values render the same title, whatever else a document says.
+        # This is stated over the declared values, so it is the part of the
+        # argument that does not depend on what any example declares. Only values
+        # declared in UNLABELLED are counted: one missing from it is reported by
+        # the check below, and naming it twice would make one defect read as two.
+        bare = sorted(
+            v
+            for v, e in axis.values.items()
+            if not e.label and not e.carried_by and (axis_name, v) in UNLABELLED
+        )
+        if len(bare) > 1:
+            problems.append(
+                f"{axis_name} values {bare} are all bare -- no label and no carried_by -- "
+                f"so their titles are identical and one would silently overwrite the other. "
+                f"Label all but one, or name the feature that carries each."
+            )
         by_label: dict[str, list[str]] = {}
         for value, entry in sorted(axis.values.items()):
             if not entry.label:
@@ -279,6 +332,19 @@ def label(axis: str, value: str | None) -> str:
     return entry.label if entry else ""
 
 
+def carried_by(axis: str, value: str | None) -> str:
+    """The feature that names a value its own label does not, or ``""``.
+
+    Like :func:`label`, an unregistered value yields ``""`` rather than raising:
+    validation is what rejects it, this only refuses to invent a qualifier.
+    """
+    if value is None:
+        return ""
+    axis_entry = AXES.get(axis)
+    entry = axis_entry.values.get(value) if axis_entry else None
+    return entry.carried_by 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 "")

unitysvc_data-0.2.13/src/unitysvc_data/examples/llm/request-template/README.md

@@ -0,0 +1,209 @@
++++
+preset_name = "llm_request_template"
+category = "request_template"
+mime_type = "json"
+file = "request-template.json"
+description = "Minimal chat request body for each request format"
+is_active = true
+is_public = true
+applies_to = { capability = "chat" }
+parameters = { version_prefix = "/v1" }
++++
+
+# llm / request-template — minimal chat payload, per request format
+
+The default request body the Test Request playground offers for a chat
+service. One document covers one capability — this one is `chat`, as its
+`applies_to` says — and from v3 it is a **list of entries**, each naming the
+request format it is written in and, from v4, the path it is sent to:
+
+```json
+[ { "format": "openai", "path_suffix": "/v1/chat/completions",
+    "body": { "max_tokens": 100, "messages": [ … ] } },
+  { "format": "anthropic", "path_suffix": "/v1/messages",
+    "body": { "max_tokens": 100, "system": "…", "messages": [ … ] } } ]
+```
+
+The playground picks the document by the service's capability and then the first
+entry whose `format` is the one the customer chose (unitysvc/unitysvc#2514).
+Entries for formats a service doesn't accept are never read. `format` names a
+request format the way the gateway does (apisix-gateways `request_meta`, which
+recognises a request as one by its path or body), and the playground compares it with the
+formats the service lists in `input_formats`, so an entry is reached only by a
+service that lists its format. Every chat service ships this one document whatever
+formats it accepts.
+
+## Other capabilities
+
+Every other capability has a sibling family, `request-template-<capability>`,
+holding the same shape for that one capability: `image-text-to-text`, `embed`,
+`rerank`, `moderate`, `image-generate`, `video-generate`, `speech-to-text` and
+`text-to-speech`. The rules they share are in
+[CONTRIBUTING](../../../../../CONTRIBUTING.md#request-templates).
+
+`image-edit` has no template. Its only code example sends `multipart/form-data`,
+which an entry's JSON body cannot express until entries can carry a
+`content_type` (unitysvc/unitysvc#2514), so there is no honest body to offer and
+a service that declares it sees the playground's empty state. The claim is checked
+against the examples (`tests/test_request_templates.py`), so it will say so when it
+stops being true.
+
+## Body (v4)
+
+| `format`           | Wire shape                                           | `path_suffix`                            |
+|--------------------|------------------------------------------------------|------------------------------------------|
+| `openai`           | OpenAI Chat Completions                              | `${__version_prefix__}/chat/completions` |
+| `anthropic`        | Anthropic Messages (`system` is top-level)           | `/v1/messages`                           |
+| `cohere`           | Cohere v2 Chat                                       | none (SDK owns the path)                 |
+| `dashscope`        | DashScope native (`input.messages` + `parameters`)   | none (bare base URL)                     |
+| `bedrock_converse` | Bedrock Converse (content blocks, `inferenceConfig`) | none (known limitation)                  |
+
+Every entry says the same thing: system prompt "You are a helpful
+assistant.", user prompt "Say hello in one sentence.", at most 100 tokens.
+
+## Path
+
+Each `path_suffix` is what a code example for that entry's format posts to,
+relative to the service's base URL. It is read off the example, and the examples
+for a format agree among themselves (the OpenAI entry's requests, shell and
+JavaScript variants all say `/chat/completions` after the prefix), rather than
+written from an API's documentation. `tests/test_request_templates.py` re-derives
+every one.
+
+`${__version_prefix__}` is the parameter the OpenAI-shaped examples use for the
+one segment a seller's upstream may move: `/v1` by default, `/compatibility/v1`
+for Cohere's compatibility surface, `/v2` for crofai, nothing for the platform's
+own facades. This family declares it with the example's default, so a listing that
+sets `params.version_prefix` for its examples gets the matching path here, and one
+that sets nothing gets `/v1`. A listing that differs in any other way can replace
+the document by title (a sibling key beside `$llm_example_collection`).
+
+`/v1/messages` carries no prefix, and that is the point rather than an omission.
+The path is a function of the capability and the format the *customer* speaks, and a
+translated format shows it: `llm_code_example_anthropic_to_openai_requests` calls
+the OpenAI upstream at `${__version_prefix__}/chat/completions` only under
+`local_testing`, while a customer calls the gateway at the literal `/v1/messages`
+and the gateway translates, whatever the upstream's layout. Only a request the
+gateway proxies without translating, an OpenAI-format one to an OpenAI-compatible
+upstream, continues into the seller's own path, which is why that one segment is a
+parameter.
+
+Three entries have no `path_suffix`, for three different reasons:
+
+- **`dashscope` posts to the bare service URL.** Every client variant of its example
+  posts to the service URL itself, so there is nothing to append and the key is
+  omitted. The page treats a missing or blank `path_suffix` as absent and appends
+  nothing, which is exactly that; writing `""` would add nothing, and `"/"` would
+  append a trailing slash, a different request.
+- **`cohere`'s example hands the base URL to the SDK**, which appends the path
+  itself, so no example states it, and the playground's own format registry already
+  carries the native Cohere chat path. The entry needs no key of its own.
+- **Known limitation: `bedrock_converse`.** Its path carries the model id, and the
+  right value varies by service inside this one entry: a native-runtime interface
+  already carries `/model/<modelId>` in its URL, so a literal `/converse` would do,
+  while a shared provider path builds `<service URL>-runtime/model/<modelId>/converse`
+  (the example has both branches). A `path_suffix` is one string per capability and
+  format, so it cannot say "it depends on the service", and the page merges `model`
+  into the body rather than substituting inside the path. That is a gap in the
+  contract (unitysvc/unitysvc#2514, noted on #2516), not an oversight and not
+  something a template can work around, so no substitution syntax is invented here.
+  The Hugging Face entries in the `embed` and `video-generate` families are the same
+  limitation.
+
+## What's intentionally missing
+
+- **No `model` field.** The gateway's routing config or the listing's
+  `upstream_access_config` selects the upstream model; the playground merges
+  the service's routing key into the body it sends. Bedrock Converse takes the
+  model id in the path instead.
+- **No `stream` field.** Non-streaming responses are simpler to
+  validate in tests; streaming variants belong in separate presets.
+- **No `bedrock_invoke` entry.** InvokeModel's body is the model family's own
+  (Anthropic-on-Bedrock, Llama, Titan, …), so there is no one body to offer. The
+  code example says as much: its body is the Anthropic Messages shape, and its
+  docstring tells the reader to adjust it for their model family. The playground
+  falls back to a blank body.
+- **No `cerebras` entry.** It is a client library, not a request format: the
+  gateway has no `cerebras` format, and `llm_code_example_cerebras` declares an
+  OpenAI upstream and calls `chat.completions.create` with `model` and
+  `messages`, the shape of the `openai` entry.
+- **One `dashscope` body, though DashScope has two chat surfaces.** The
+  gateway has a single `dashscope` format, and an entry is unique per format, so
+  only one body fits. This is the text-generation shape (`content` is a string,
+  `result_format = "message"`), matching `llm_code_example_chat_dashscope_text_*`.
+  The multimodal-generation surface (`llm_code_example_chat_dashscope_*`) takes
+  `content` as a list of `{"text": …}` parts and no `result_format`, so a service
+  on it is offered a body that is not its own and must edit it.
+
+## Conventions
+
+- `max_tokens` is kept small (≤ 100) so the request completes fast
+  against any upstream regardless of per-token latency.
+- A new request format gets a new entry here, in a new version.
+- Where a body goes beyond what its code example sends, it is carried over
+  exactly as v2 published it: the `cohere` entry's `max_tokens`, the
+  `bedrock_converse` entry's `system`, and the `dashscope` entry's system message
+  are not in any example, so nothing the test runner executes exercises them.
+
+## Versions
+
+### v4 — a path for each entry
+
+- Entries gained `path_suffix`: `openai` and `anthropic` have one, read off their
+  examples; `cohere`, `dashscope` and `bedrock_converse` do not, each for the reason
+  under Path. The bodies, the formats and their order are v3's, and a test pins that.
+- **A new version, not an amendment of v3.** CONTRIBUTING's "one exception" needs
+  the old behaviour to be gone, and chat's is not: the playground falls back to the
+  format's own default path for a chat request, so a template without paths works on
+  every service whose OpenAI surface is at `/v1`. v3 stays exactly as published, with
+  no paths, for a listing that pins it. (The eight sibling families were amended in
+  place instead: a template for a non-chat capability without a path cannot be sent,
+  so their v1 had nothing working to pin to.)
+- The front-matter gained `parameters = { version_prefix = "/v1" }`. It is shared by
+  every version of the family, but v1 to v3 contain no placeholder, so their content
+  is byte-identical; only the record's `file_path` now points at a per-process copy,
+  as for any preset that declares parameters.
+- `llm_request_template` resolves to v4. Needs nothing of the page: it ignores keys
+  it does not read, so data and page can ship in either order, and a page that reads
+  `path_suffix` falls back to today's behaviour for an entry that has none.
+
+### v3 — an entry list (superseded by v4)
+
+- A list of `{"format", "body"}` entries instead of a dict keyed by format. The
+  bodies are v2's, key for key and in the same order; only the packaging moved.
+  A dict makes the format the key and leaves nothing to say about a document
+  holding more than one capability; entries name their own format, and the
+  playground's reader (unitysvc/unitysvc#2515) takes the first entry that
+  matches, so a repeated format would be silently unreachable — which
+  `tools/build.py` now refuses.
+- The title, `applies_to` and preset name are unchanged, so this is the same
+  document on every published service: `llm_request_template` resolved to v3 in
+  0.2.12 (it resolves to v4 from 0.2.13) and re-uploading a catalog replaces the body
+  in place.
+- Needs a playground that reads entry lists (unitysvc/unitysvc#2515). That page
+  reads both shapes, so data and page can ship in either order **only because
+  of that** — an older page shows a list as the raw request body.
+
+### v2 — one body per request format (superseded by v3)
+
+- Keyed by format: `openai`, `anthropic`, `cohere`, `dashscope`,
+  `bedrock_converse`. The `openai` entry is v1's body.
+- `applies_to` no longer requires an OpenAI upstream, so every chat service gets
+  this template whatever its dialect — which is what makes one document the
+  right shape: a service accepting both OpenAI and Anthropic needs ONE default
+  body, and the playground indexes it by the format the customer picked.
+  `llm_request_template_anthropic` is removed rather than kept alongside; per-format
+  templates would give such a service two `Default request body` documents and no
+  way to say which is the default.
+- Needs a playground that picks a format's entry (unitysvc/unitysvc#2508).
+  An older one shows the whole map as the body.
+- **Retired, not removed.** Versions are append-only, so the file stays and
+  `llm_request_template_v2` still resolves for a listing that pins it; none of
+  the `unitysvc-services-*` repos does. The version-less alias has moved on (v3,
+  now v4), which is all that `llm_example_collection` ever resolves, so nothing
+  generated serves v2.
+
+### v1 — initial release
+
+- Two messages (system + user), `max_tokens = 100`, no model field,
+  non-streaming.

unitysvc_data-0.2.13/src/unitysvc_data/examples/llm/request-template/request-template-v3.json

@@ -0,0 +1,91 @@
+[
+  {
+    "format": "openai",
+    "body": {
+      "max_tokens": 100,
+      "messages": [
+        {
+          "role": "system",
+          "content": "You are a helpful assistant."
+        },
+        {
+          "role": "user",
+          "content": "Say hello in one sentence."
+        }
+      ]
+    }
+  },
+  {
+    "format": "anthropic",
+    "body": {
+      "max_tokens": 100,
+      "system": "You are a helpful assistant.",
+      "messages": [
+        {
+          "role": "user",
+          "content": "Say hello in one sentence."
+        }
+      ]
+    }
+  },
+  {
+    "format": "cohere",
+    "body": {
+      "max_tokens": 100,
+      "messages": [
+        {
+          "role": "system",
+          "content": "You are a helpful assistant."
+        },
+        {
+          "role": "user",
+          "content": "Say hello in one sentence."
+        }
+      ]
+    }
+  },
+  {
+    "format": "dashscope",
+    "body": {
+      "input": {
+        "messages": [
+          {
+            "role": "system",
+            "content": "You are a helpful assistant."
+          },
+          {
+            "role": "user",
+            "content": "Say hello in one sentence."
+          }
+        ]
+      },
+      "parameters": {
+        "max_tokens": 100,
+        "result_format": "message"
+      }
+    }
+  },
+  {
+    "format": "bedrock_converse",
+    "body": {
+      "system": [
+        {
+          "text": "You are a helpful assistant."
+        }
+      ],
+      "messages": [
+        {
+          "role": "user",
+          "content": [
+            {
+              "text": "Say hello in one sentence."
+            }
+          ]
+        }
+      ],
+      "inferenceConfig": {
+        "maxTokens": 100
+      }
+    }
+  }
+]