unitysvc-data 0.1.38__tar.gz → 0.1.40__tar.gz

{unitysvc_data-0.1.38/src/unitysvc_data.egg-info → unitysvc_data-0.1.40}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: unitysvc-data
-Version: 0.1.38
+Version: 0.1.40
 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-0.1.38 → unitysvc_data-0.1.40}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "unitysvc-data"
-version = "0.1.38"
+version = "0.1.40"
 description = "Standard examples and presets for UnitySVC data packages"
 readme = "README.md"
 authors = [{ name = "Bo Peng", email = "bo.peng@unitysvc.com" }]

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/_manifest.json

@@ -61,14 +61,15 @@
     "llm_code_example_ttv_javascript": "llm_code_example_ttv_javascript_v1",
     "llm_code_example_ttv_requests": "llm_code_example_ttv_requests_v1",
     "llm_code_example_ttv_shell": "llm_code_example_ttv_shell_v2",
-    "llm_code_example_vision_javascript": "llm_code_example_vision_javascript_v1",
-    "llm_code_example_vision_requests": "llm_code_example_vision_requests_v1",
-    "llm_code_example_vision_shell": "llm_code_example_vision_shell_v2",
+    "llm_code_example_vision_javascript": "llm_code_example_vision_javascript_v2",
+    "llm_code_example_vision_requests": "llm_code_example_vision_requests_v2",
+    "llm_code_example_vision_shell": "llm_code_example_vision_shell_v3",
     "llm_connectivity": "llm_connectivity_v1",
     "llm_connectivity_anthropic": "llm_connectivity_anthropic_v1",
     "llm_connectivity_embed": "llm_connectivity_embed_v1",
     "llm_connectivity_embed_image": "llm_connectivity_embed_image_v1",
     "llm_connectivity_transcription": "llm_connectivity_transcription_v1",
+    "llm_connectivity_tts": "llm_connectivity_tts_v1",
     "llm_description": "llm_description_v2",
     "llm_request_template": "llm_request_template_v1",
     "llm_request_template_anthropic": "llm_request_template_anthropic_v1",
@@ -2122,7 +2123,7 @@
     },
     "llm_code_example_transcription_javascript_v1": {
       "applies_to": {
-        "capability": "speech-transcribe"
+        "capability": "speech-to-text"
       },
       "category": "code_example",
       "description": "JavaScript example: transcribe audio via OpenAI-compatible /audio/transcriptions",
@@ -2143,7 +2144,7 @@
     },
     "llm_code_example_transcription_requests_v1": {
       "applies_to": {
-        "capability": "speech-transcribe"
+        "capability": "speech-to-text"
       },
       "category": "code_example",
       "description": "Python example: transcribe a pre-recorded audio file with an OpenAI-compatible audio transcription endpoint",
@@ -2167,7 +2168,7 @@
     },
     "llm_code_example_transcription_shell_v1": {
       "applies_to": {
-        "capability": "speech-transcribe"
+        "capability": "speech-to-text"
       },
       "category": "code_example",
       "description": "Bash example: transcribe audio via OpenAI-compatible /audio/transcriptions using curl",
@@ -2188,7 +2189,7 @@
     },
     "llm_code_example_transcription_shell_v2": {
       "applies_to": {
-        "capability": "speech-transcribe"
+        "capability": "speech-to-text"
       },
       "category": "code_example",
       "description": "Bash example: transcribe audio via OpenAI-compatible /audio/transcriptions using curl",
@@ -2210,7 +2211,7 @@
     },
     "llm_code_example_tts_javascript_v1": {
       "applies_to": {
-        "capability": "speech-synthesize"
+        "capability": "text-to-speech"
       },
       "category": "code_example",
       "description": "JavaScript example: synthesize speech via OpenAI-compatible /audio/speech",
@@ -2230,7 +2231,7 @@
     },
     "llm_code_example_tts_requests_v1": {
       "applies_to": {
-        "capability": "speech-synthesize"
+        "capability": "text-to-speech"
       },
       "category": "code_example",
       "description": "Python example: synthesize speech from text using an OpenAI-compatible audio/speech endpoint",
@@ -2253,7 +2254,7 @@
     },
     "llm_code_example_tts_shell_v1": {
       "applies_to": {
-        "capability": "speech-synthesize"
+        "capability": "text-to-speech"
       },
       "category": "code_example",
       "description": "Bash example: synthesize speech via OpenAI-compatible /audio/speech using curl",
@@ -2357,7 +2358,7 @@
     },
     "llm_code_example_vision_javascript_v1": {
       "applies_to": {
-        "capability": "chat",
+        "capability": "image-text-to-text",
         "dialect": "openai",
         "feature": "vision",
         "upstream": "openai"
@@ -2378,9 +2379,32 @@
       "source_readme": "llm/code-example-vision-javascript/README.md",
       "version": 1
     },
+    "llm_code_example_vision_javascript_v2": {
+      "applies_to": {
+        "capability": "image-text-to-text",
+        "dialect": "openai",
+        "feature": "vision",
+        "upstream": "openai"
+      },
+      "category": "code_example",
+      "description": "JavaScript example: describe an image via OpenAI-compatible /chat/completions vision messages",
+      "example_file": "llm/code-example-vision-javascript/code-example-vision-v2.js.j2",
+      "is_active": true,
+      "is_public": true,
+      "meta": {
+        "variant": "Vision"
+      },
+      "mime_type": "javascript",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_code_example_vision_javascript",
+      "source_readme": "llm/code-example-vision-javascript/README.md",
+      "version": 2
+    },
     "llm_code_example_vision_requests_v1": {
       "applies_to": {
-        "capability": "chat",
+        "capability": "image-text-to-text",
         "dialect": "openai",
         "feature": "vision",
         "upstream": "openai"
@@ -2404,9 +2428,35 @@
       "source_readme": "llm/code-example-vision-requests/README.md",
       "version": 1
     },
+    "llm_code_example_vision_requests_v2": {
+      "applies_to": {
+        "capability": "image-text-to-text",
+        "dialect": "openai",
+        "feature": "vision",
+        "upstream": "openai"
+      },
+      "category": "code_example",
+      "description": "Python example: ask a vision-capable LLM about a publicly-hosted image",
+      "example_file": "llm/code-example-vision-requests/code-example-vision-v2.py.j2",
+      "is_active": true,
+      "is_public": true,
+      "meta": {
+        "requirements": [
+          "requests"
+        ],
+        "variant": "Vision"
+      },
+      "mime_type": "python",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_code_example_vision_requests",
+      "source_readme": "llm/code-example-vision-requests/README.md",
+      "version": 2
+    },
     "llm_code_example_vision_shell_v1": {
       "applies_to": {
-        "capability": "chat",
+        "capability": "image-text-to-text",
         "dialect": "openai",
         "feature": "vision",
         "upstream": "openai"
@@ -2429,7 +2479,7 @@
     },
     "llm_code_example_vision_shell_v2": {
       "applies_to": {
-        "capability": "chat",
+        "capability": "image-text-to-text",
         "dialect": "openai",
         "feature": "vision",
         "upstream": "openai"
@@ -2451,6 +2501,30 @@
       "source_readme": "llm/code-example-vision-shell/README.md",
       "version": 2
     },
+    "llm_code_example_vision_shell_v3": {
+      "applies_to": {
+        "capability": "image-text-to-text",
+        "dialect": "openai",
+        "feature": "vision",
+        "upstream": "openai"
+      },
+      "category": "code_example",
+      "description": "Bash example: describe an image via OpenAI-compatible /chat/completions using curl",
+      "example_file": "llm/code-example-vision-shell/code-example-vision-v3.sh.j2",
+      "is_active": true,
+      "is_public": true,
+      "meta": {
+        "output_contains": "example ok",
+        "variant": "Vision"
+      },
+      "mime_type": "bash",
+      "parameters": {
+        "version_prefix": "/v1"
+      },
+      "preset_name": "llm_code_example_vision_shell",
+      "source_readme": "llm/code-example-vision-shell/README.md",
+      "version": 3
+    },
     "llm_connectivity_anthropic_v1": {
       "applies_to": {
         "capability": "chat",
@@ -2513,7 +2587,7 @@
     },
     "llm_connectivity_transcription_v1": {
       "applies_to": {
-        "capability": "speech-transcribe"
+        "capability": "speech-to-text"
       },
       "category": "connectivity_test",
       "description": "Verify an OpenAI-compatible audio-transcription endpoint by sending a tiny known audio file",
@@ -2532,6 +2606,29 @@
       "source_readme": "llm/connectivity-transcription/README.md",
       "version": 1
     },
+    "llm_connectivity_tts_v1": {
+      "applies_to": {
+        "capability": "text-to-speech"
+      },
+      "category": "connectivity_test",
+      "description": "Verify an OpenAI-compatible text-to-speech endpoint by synthesizing a short phrase and asserting real audio comes back",
+      "example_file": "llm/connectivity-tts/connectivity-tts-v1.sh.j2",
+      "is_active": true,
+      "is_public": false,
+      "meta": {
+        "output_contains": "connectivity ok"
+      },
+      "mime_type": "bash",
+      "parameters": {
+        "min_bytes": "1024",
+        "response_format": "wav",
+        "version_prefix": "/v1",
+        "voice": "alloy"
+      },
+      "preset_name": "llm_connectivity_tts",
+      "source_readme": "llm/connectivity-tts/README.md",
+      "version": 1
+    },
     "llm_connectivity_v1": {
       "applies_to": {
         "capability": "chat",

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-transcription-javascript/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Transcription" }
 parameters = { version_prefix = "/v1", language = "en" }
-applies_to = { capability = "speech-transcribe" }
+applies_to = { capability = "speech-to-text" }
 +++
 
 # llm / code-example-transcription-javascript — audio transcription via `fetch`

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-transcription-requests/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Transcription", requirements = ["requests"] }
 parameters = { version_prefix = "/v1", language = "en" }
-applies_to = { capability = "speech-transcribe" }
+applies_to = { capability = "speech-to-text" }
 +++
 
 # llm / code-example-transcription-requests — audio transcription via `requests`

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-transcription-shell/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Transcription", output_contains = "example ok" }
 parameters = { version_prefix = "/v1", language = "en" }
-applies_to = { capability = "speech-transcribe" }
+applies_to = { capability = "speech-to-text" }
 
 [versions.v1]
 # v1 predates the response-shape assertion and prints no sentinel.

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-tts-javascript/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Text to speech" }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "speech-synthesize" }
+applies_to = { capability = "text-to-speech" }
 +++
 
 # llm / code-example-tts-javascript — text-to-speech via `fetch`

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-tts-requests/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Text to speech", requirements = ["requests"] }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "speech-synthesize" }
+applies_to = { capability = "text-to-speech" }
 +++
 
 # llm / code-example-tts-requests — text-to-speech via `requests`

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-tts-shell/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Text to speech" }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "speech-synthesize" }
+applies_to = { capability = "text-to-speech" }
 +++
 
 # llm / code-example-tts-shell — text-to-speech via `curl`

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-vision-javascript/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Vision" }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "chat", dialect = "openai", upstream = "openai", feature = "vision" }
+applies_to = { capability = "image-text-to-text", dialect = "openai", upstream = "openai", feature = "vision" }
 +++
 
 # llm / code-example-vision-javascript — vision via `fetch`
@@ -34,3 +34,17 @@ Optional:
 ## Versions
 
 ### v1 — initial release
+
+### v2 — inline the image, and assert the reply shape
+
+- Fetches `IMAGE_URL` and sends it as a `data:` URI, so the request no
+  longer depends on the provider being able to reach that host.
+  Server-side fetching was intermittently failing (observed on DeepSeek:
+  3 of 4 identical calls succeeded, the fourth returned
+  `400 Failed to download image`), which made every vision example a
+  coin-flip in CI for reasons unrelated to the service.
+- Takes the media type from the image response's `content-type` rather
+  than assuming JPEG.
+- Adds the `choices` assertion and the `example ok` sentinel the family
+  metadata already expects — a 200 carrying an error object no longer
+  reads as a pass.

unitysvc_data-0.1.40/src/unitysvc_data/examples/llm/code-example-vision-javascript/code-example-vision-v2.js.j2

@@ -0,0 +1,61 @@
+async function main() {
+  const imageUrl =
+    process.env.IMAGE_URL ||
+    "https://raw.githubusercontent.com/unitysvc/unitysvc-data/main/src/unitysvc_data/examples/llm/test-image.jpg";
+  const prompt = process.env.PROMPT || "Describe this image.";
+
+  // Fetch the image HERE and inline it as a data URI, rather than handing the
+  // model a URL and letting it fetch the bytes itself. Server-side fetching is
+  // the provider's network, not yours: some cannot reach an arbitrary host at
+  // all, and others manage it only most of the time, so the call fails with a
+  // download error that has nothing to do with your request. Inlining removes
+  // that dependency — the only fetch is this one, and it either works or fails
+  // where you can see it.
+  //
+  // To send your own local image instead, drop the fetch and read the file:
+  //   const imageB64 = require("fs").readFileSync("my-image.jpg").toString("base64");
+  const imageResponse = await fetch(imageUrl);
+  if (!imageResponse.ok) {
+    throw new Error(`HTTP ${imageResponse.status} fetching ${imageUrl}`);
+  }
+  const mediaType = imageResponse.headers.get("content-type") || "image/jpeg";
+  const imageB64 = Buffer.from(await imageResponse.arrayBuffer()).toString("base64");
+
+  const response = await fetch("{{ service_base_url }}${__version_prefix__}/chat/completions", {
+    method: "POST",
+    headers: {
+      Authorization: `Bearer ${process.env.UNITYSVC_API_KEY}`,
+      "Content-Type": "application/json",
+    },
+    body: JSON.stringify({
+      model: "{{ routing_key.model }}",
+      messages: [
+        {
+          role: "user",
+          content: [
+            { type: "text", text: prompt },
+            { type: "image_url", image_url: { url: `data:${mediaType};base64,${imageB64}` } },
+          ],
+        },
+      ],
+    }),
+  });
+  if (!response.ok) {
+    throw new Error(`HTTP ${response.status}: ${await response.text()}`);
+  }
+
+  // A 2xx does not mean the call did what you asked: an error object, or a
+  // reply from a different capability, also returns 200. Assert the reply
+  // really is a chat completion before treating this as success.
+  const data = await response.json();
+  if (!data.choices) {
+    throw new Error(`unexpected response: no "choices" in the reply: ${JSON.stringify(data)}`);
+  }
+  console.log(data.choices[0].message.content);
+  console.log("example ok");
+}
+
+main().catch((err) => {
+  console.error(err.message);
+  process.exit(1);
+});

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-vision-requests/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Vision", requirements = ["requests"] }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "chat", dialect = "openai", upstream = "openai", feature = "vision" }
+applies_to = { capability = "image-text-to-text", dialect = "openai", upstream = "openai", feature = "vision" }
 +++
 
 # llm / code-example-vision-requests — image understanding via `requests`
@@ -54,3 +54,17 @@ Optional:
   image_url referencing a public Wikimedia Commons photo by
   default).
 - `response.raise_for_status()` so non-2xx exits non-zero.
+
+### v2 — inline the image, and assert the reply shape
+
+- Downloads `IMAGE_URL` and sends it as a `data:` URI, so the request no
+  longer depends on the provider being able to reach that host.
+  Server-side fetching was intermittently failing (observed on DeepSeek:
+  3 of 4 identical calls succeeded, the fourth returned
+  `400 Failed to download image`), which made every vision example a
+  coin-flip in CI for reasons unrelated to the service.
+- Takes the media type from the image response's `Content-Type` rather
+  than assuming JPEG.
+- Adds the `"choices"` assertion and the `example ok` sentinel the family
+  metadata already expects — a 200 carrying an error object no longer
+  reads as a pass.

unitysvc_data-0.1.40/src/unitysvc_data/examples/llm/code-example-vision-requests/code-example-vision-v2.py.j2

@@ -0,0 +1,60 @@
+import base64
+import os
+
+import requests
+
+image_url = os.environ.get(
+    "IMAGE_URL",
+    "https://raw.githubusercontent.com/unitysvc/unitysvc-data/main/src/unitysvc_data/examples/llm/test-image.jpg",
+)
+prompt = os.environ.get("PROMPT", "Describe this image.")
+
+# Fetch the image HERE and inline it as a data URI, rather than handing the
+# model a URL and letting it fetch the bytes itself. Server-side fetching is
+# the provider's network, not yours: some cannot reach an arbitrary host at
+# all, and others manage it only most of the time, so the call fails with a
+# download error that has nothing to do with your request. Inlining removes
+# that dependency — the only fetch is this one, and it either works or fails
+# where you can see it.
+#
+# To send your own local image instead, drop the request and read the file:
+#   media_type, image_bytes = "image/jpeg", open("my-image.jpg", "rb").read()
+image_response = requests.get(image_url, timeout=30)
+if image_response.status_code >= 400:
+    raise RuntimeError(f"HTTP {image_response.status_code} fetching {image_url}: {image_response.text[:200]}")
+media_type = image_response.headers.get("Content-Type", "image/jpeg")
+image_b64 = base64.b64encode(image_response.content).decode()
+
+response = requests.post(
+    "{{ service_base_url }}${__version_prefix__}/chat/completions",
+    headers={
+        "Authorization": f"Bearer {os.environ['UNITYSVC_API_KEY']}",
+        "Content-Type": "application/json",
+    },
+    json={
+        "model": "{{ routing_key.model }}",
+        "messages": [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": prompt},
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": f"data:{media_type};base64,{image_b64}"},
+                    },
+                ],
+            }
+        ],
+    },
+)
+if response.status_code >= 400:
+    raise RuntimeError(f"HTTP {response.status_code} from {response.url}: {response.text}")
+
+# A 2xx does not mean the call did what you asked: an error object, or a
+# reply from a different capability, also returns 200. Assert the reply
+# really is a chat completion before treating this as success.
+data = response.json()
+if "choices" not in data:
+    raise RuntimeError(f'unexpected response: no "choices" in the reply: {data}')
+print(data["choices"][0]["message"]["content"])
+print("example ok")

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/code-example-vision-shell/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = true
 meta = { variant = "Vision", output_contains = "example ok" }
 parameters = { version_prefix = "/v1" }
-applies_to = { capability = "chat", dialect = "openai", upstream = "openai", feature = "vision" }
+applies_to = { capability = "image-text-to-text", dialect = "openai", upstream = "openai", feature = "vision" }
 
 [versions.v1]
 # v1 predates the response-shape assertion and prints no sentinel.
@@ -37,3 +37,15 @@ Optional:
 ## Versions
 
 ### v1 — initial release
+
+### v3 — inline the image instead of asking the model to fetch it
+
+- Downloads `IMAGE_URL` with `curl` and sends it as a `data:` URI, so the
+  request no longer depends on the provider being able to reach that host.
+  Server-side fetching was intermittently failing (observed on DeepSeek:
+  3 of 4 identical calls succeeded, the fourth returned
+  `400 Failed to download image`), which made every vision example a
+  coin-flip in CI for reasons unrelated to the service.
+- Takes the media type from the response's `Content-Type` rather than
+  assuming JPEG, so swapping in a PNG needs no edit.
+- `base64` wraps output on GNU but not BSD; `tr -d '\n'` normalises both.

unitysvc_data-0.1.40/src/unitysvc_data/examples/llm/code-example-vision-shell/code-example-vision-v3.sh.j2

@@ -0,0 +1,50 @@
+#!/bin/bash
+set -e -o pipefail
+
+: "${UNITYSVC_API_KEY:?UNITYSVC_API_KEY is not set}"
+
+image_url="${IMAGE_URL:-https://raw.githubusercontent.com/unitysvc/unitysvc-data/main/src/unitysvc_data/examples/llm/test-image.jpg}"
+prompt="${PROMPT:-Describe this image.}"
+
+# Fetch the image HERE and inline it as a data URI, rather than handing the
+# model a URL and letting it fetch the bytes itself. Server-side fetching is
+# the provider's network, not yours: some cannot reach an arbitrary host at
+# all, and others manage it only most of the time, so the call fails with a
+# download error that has nothing to do with your request. Inlining removes
+# that dependency — the only fetch is this one, and it either works or fails
+# where you can see it.
+#
+# To send your own local image instead, drop the curl and read the file:
+#   media_type="image/jpeg"; image_b64=$(base64 < ./my-image.jpg | tr -d '\n')
+tmp_image=$(mktemp)
+trap 'rm -f "$tmp_image"' EXIT
+media_type=$(curl --fail-with-body -sS -o "$tmp_image" -w '%{content_type}' "$image_url")
+# `base64` wraps output on GNU but not BSD; strip newlines so both produce one line.
+image_b64=$(base64 < "$tmp_image" | tr -d '\n')
+
+response=$(curl --fail-with-body -sS "{{ service_base_url }}${__version_prefix__}/chat/completions" \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer ${UNITYSVC_API_KEY}" \
+  -d "{
+    \"model\": \"{{ routing_key.model }}\",
+    \"messages\": [
+      {
+        \"role\": \"user\",
+        \"content\": [
+          {\"type\": \"text\", \"text\": \"${prompt}\"},
+          {\"type\": \"image_url\", \"image_url\": {\"url\": \"data:${media_type};base64,${image_b64}\"}}
+        ]
+      }
+    ]
+  }")
+
+echo "$response"
+
+# A 2xx does not mean the call did what you asked: an error object, or a
+# reply from a different capability, also returns 200. Assert the reply
+# really is a chat completion before treating this as success.
+if ! printf '%s' "$response" | grep -q '"choices"'; then
+  echo 'unexpected response: no "choices" in the reply' >&2
+  exit 1
+fi
+echo "example ok"

{unitysvc_data-0.1.38 → unitysvc_data-0.1.40}/src/unitysvc_data/examples/llm/connectivity-transcription/README.md

@@ -8,7 +8,7 @@ is_active = true
 is_public = false
 meta = { output_contains = "connectivity ok" }
 parameters = { version_prefix = "/v1", language = "en" }
-applies_to = { capability = "speech-transcribe" }
+applies_to = { capability = "speech-to-text" }
 +++
 
 # llm / connectivity-transcription — audio transcription smoke test

unitysvc_data-0.1.40/src/unitysvc_data/examples/llm/connectivity-tts/README.md

@@ -0,0 +1,80 @@
++++
+preset_name = "llm_connectivity_tts"
+category = "connectivity_test"
+mime_type = "bash"
+file = "connectivity-tts.sh.j2"
+description = "Verify an OpenAI-compatible text-to-speech endpoint by synthesizing a short phrase and asserting real audio comes back"
+is_active = true
+is_public = false
+meta = { output_contains = "connectivity ok" }
+parameters = { version_prefix = "/v1", voice = "alloy", response_format = "wav", min_bytes = "1024" }
+applies_to = { capability = "text-to-speech" }
++++
+
+# llm / connectivity-tts — text-to-speech smoke test
+
+Connectivity check for OpenAI-compatible **text-to-speech** services routed
+through the UnitySVC LLM gateway. The standard `llm_connectivity` preset POSTs
+to `/chat/completions` and is rejected by TTS-only models (they expose
+`/audio/speech`, not the chat surface), which fails the whole service at review
+with an upstream 4xx.
+
+This closes the gap recorded in unitysvc/unitysvc#1781: TTS shipped code
+examples but no connectivity probe, and connectivity is mandatory to activate —
+so TTS services could never leave `draft` however complete their examples were.
+
+## Asserting on audio, not JSON
+
+Every other probe in this family greps the response body for a field —
+`"choices"` for chat, `"text"` for transcription. A successful TTS response is
+**audio bytes**, so there is nothing to grep. Three checks stand in for it:
+
+1. **HTTP 2xx.**
+2. **At least `min_bytes` of body.** A stub, an empty file or a short error page
+   fails. 1 KiB is comfortably below a second of any real codec output and
+   comfortably above an error envelope.
+3. **The body does not start with `{`.** Some gateways answer 200 with a JSON
+   error envelope; byte count alone would accept a long one. Only the first byte
+   is inspected, so audio that happens to contain a brace later still passes.
+
+## Template variables (filled in by the platform when rendering for a given access interface)
+
+- `{{ service_base_url }}` — endpoint base URL.
+- `{{ routing_key.model }}` — model id.
+
+## Parameters (per-listing)
+
+- `version_prefix` (default `/v1`) — path segment between `service_base_url`
+  and `/audio/speech`. Override to e.g. `/compatibility/v1` for providers that
+  expose the OpenAI-compat shape under a non-standard prefix.
+- `voice` (default `alloy`) — voice id. Providers do not share a voice
+  vocabulary, so a service whose upstream rejects `alloy` must set its own
+  (groq's `canopylabs/orpheus-*` and parasail's `resemble-tts-en` each have
+  their own list).
+- `response_format` (default `wav`) — container requested from the upstream.
+- `min_bytes` (default `1024`) — floor for the "this is real audio" assertion.
+
+## Environment variables (read at runtime)
+
+- `UNITYSVC_API_KEY` — required.
+
+## Route status
+
+`/v1/audio/speech` is **unverified through the gateway** — no catalog exercises
+it yet, unlike `/v1/audio/transcriptions`, which cohere ships today. This probe
+is how that gets settled: running it against a real TTS service either passes,
+proving the route, or produces a concrete gateway bug to file. See
+unitysvc/unitysvc#1781 blocker 2.
+
+## Versions
+
+### v1 — initial release
+
+- POST `/audio/speech` with model, a fixed short input, voice and
+  `response_format`.
+- 2xx, byte-count and not-JSON assertions as described above.
+- Removes the downloaded response before exiting so repeated runs do not
+  accumulate audio in the temp dir.
+- Output contains `connectivity ok` — paired with the
+  `output_contains = "connectivity ok"` meta so the run-tests flow can confirm
+  a real round-trip.