Slim Azure-only llm clients; remove description/preamble

description.md

@@ -1,27 +0,0 @@
-## How it works
-
-<p style="background-color: #fff9f9; border: 1px solid #ff0000; padding: 10px;">
-    Warning: This demo calls a live LLM API and is experimental.
-</p>
-
-Unlike embedding-based approaches that score token similarity directly, this demo compares two **generate-then-label** pipelines on the same input pair:
-
-### Auxiliary synchronization
-
-1. **Edit** — The LLM minimally edits Text A so it stays in A's language but becomes semantically closer to Text B (and vice versa).
-2. **Diff** — A sequence alignment between the original and edited text yields binary labels per token.
-3. **Highlight** — Edited tokens are shown in orange.
-
-### Span marking
-
-1. **Mark** — The LLM wraps dissimilar spans in Text A with `{{ }}` markers (without rewriting), and vice versa.
-2. **Parse** — Markers are mapped onto tokens to yield binary labels.
-3. **Highlight** — Marked tokens are shown in orange.
-
-The inputs may be in different languages. When no difference is needed, the model responds with `pass`.
-
-**Model:** `gpt-5.6-terra` (Azure, no reasoning) with prompt configs `config.v0.6.json` (auxiliary synchronization) and `config.span_marking.v0.6.json` (span marking).
-
-More resources:
-- Dataset: [ZurichNLP/SwissGov-RSD](https://huggingface.co/datasets/ZurichNLP/SwissGov-RSD)
-- Related work: [Unsupervised Semantic Diff](https://huggingface.co/spaces/ZurichNLP/unsupervised-semantic-diff) (embedding-based)

llm_query/llm_client.py

@@ -23,17 +23,17 @@ class LLMResponse:
     response_obj: dict
 
 
-class LLMClient:
+class AzureClient:
 
-    BASE_URL = "https://api.swissai.svc.cscs.ch/v1"
-    API_KEY_NAME = "CSCS_SERVING_API"
+    API_KEY_NAME = "AZURE_API_KEY"
 
-    def __init__(self,
-                 model_name: str = "google/gemma-4-31B-it",
-                 config_name: str = "config.v0.json",
-                 cache_directory: Optional[Path] = None,
-                 use_cache: bool = True,
-                 ):
+    def __init__(
+            self,
+            model_name: str = "gpt-5.6-terra",
+            config_name: str = "config.v0.6.json",
+            cache_directory: Optional[Path] = None,
+            use_cache: bool = True,
+    ):
         self.config_name = config_name
         with open(Path(__file__).parent / "configs" / config_name) as f:
             self.config = json.load(f)
@@ -45,6 +45,7 @@ class LLMClient:
                 assert "edited_text_a" in example
 
         self.model_name = model_name
+        self.BASE_URL = os.environ.get("AZURE_BASE_URL")
         self.cache_directory = cache_directory or Path(__file__).parent / ".llm_cache"
         self.client = openai.Client(
             api_key=os.environ.get(self.API_KEY_NAME),
@@ -65,7 +66,7 @@ class LLMClient:
         return (self.model_name, self.config_name, text_a, text_b)
 
     def _completion_extra_kwargs(self) -> dict:
-        kwargs = {}
+        kwargs = {"reasoning_effort": "none"}
         for key in ("seed", "temperature"):
             if key in self.config:
                 kwargs[key] = self.config[key]
@@ -95,10 +96,7 @@ class LLMClient:
         })
         return messages
 
-    def query(self,
-              text_a: str,
-              text_b: str,
-              ) -> LLMResponse:
+    def query(self, text_a: str, text_b: str) -> LLMResponse:
         cache_key = self._get_cache_key(text_a, text_b)
 
         if self.cache is not None and cache_key in self.cache:
@@ -143,55 +141,3 @@ class LLMClient:
             return llm_response
         llm_response.edited_text_a = self._normalize_edited_text(content, text_a)
         return llm_response
-
-
-class GPTClient(LLMClient):
-
-    BASE_URL = None  # Default
-    API_KEY_NAME = "OPENAI_API_KEY"
-
-    def _completion_extra_kwargs(self) -> dict:
-        return {
-            "reasoning_effort": "none",
-            **super()._completion_extra_kwargs(),
-        }
-
-
-class GeminiClient(LLMClient):
-
-    BASE_URL = "http://172.23.205.120:4000/v1"
-    API_KEY_NAME = "LITELLM_API_KEY"
-
-    def _completion_extra_kwargs(self) -> dict:
-        return {
-            "reasoning_effort": "minimal",
-            **super()._completion_extra_kwargs(),
-        }
-
-
-class AzureClient(LLMClient):
-
-    API_KEY_NAME = "AZURE_API_KEY"
-
-    def __init__(self, *args, **kwargs):
-        self.BASE_URL = os.environ.get("AZURE_BASE_URL")
-        super().__init__(*args, **kwargs)
-
-    def _completion_extra_kwargs(self) -> dict:
-        return {
-            "reasoning_effort": "none",
-            **super()._completion_extra_kwargs(),
-        }
-
-
-class DeepseekClient(LLMClient):
-
-    BASE_URL = "https://api.deepseek.com"
-    API_KEY_NAME = "DEEPSEEK_API_KEY"
-
-    def _completion_extra_kwargs(self) -> dict:
-        return {
-            "reasoning_effort": "high",
-            "extra_body": {"thinking": {"type": "enabled"}},
-            **super()._completion_extra_kwargs(),
-        }

llm_query/span_marking/llm_client.py

@@ -26,14 +26,13 @@ class SpanMarkingLLMResponse:
     response_obj: dict
 
 
-class SpanMarkingLLMClient:
+class AzureClient:
 
-    BASE_URL = "https://api.swissai.svc.cscs.ch/v1"
-    API_KEY_NAME = "CSCS_SERVING_API"
+    API_KEY_NAME = "AZURE_API_KEY"
 
     def __init__(
             self,
-            model_name: str = "google/gemma-4-31B-it",
+            model_name: str = "gpt-5.6-terra",
             config_name: str = "config.span_marking.v0.6.json",
             cache_directory: Optional[Path] = None,
             use_cache: bool = True,
@@ -49,6 +48,7 @@ class SpanMarkingLLMClient:
                 assert "marked_text_a" in example
 
         self.model_name = model_name
+        self.BASE_URL = os.environ.get("AZURE_BASE_URL")
         self.cache_directory = cache_directory or Path(__file__).parent / ".llm_cache"
         self.client = openai.Client(
             api_key=os.environ.get(self.API_KEY_NAME),
@@ -69,7 +69,7 @@ class SpanMarkingLLMClient:
         return (self.model_name, self.config_name, text_a, text_b)
 
     def _completion_extra_kwargs(self) -> dict:
-        kwargs = {}
+        kwargs = {"reasoning_effort": "none"}
         for key in ("seed", "temperature"):
             if key in self.config:
                 kwargs[key] = self.config[key]
@@ -151,55 +151,3 @@ class SpanMarkingLLMClient:
             return llm_response
         llm_response.marked_text_a = self._normalize_marked_text(content, text_a)
         return llm_response
-
-
-class GPTClient(SpanMarkingLLMClient):
-
-    BASE_URL = None
-    API_KEY_NAME = "OPENAI_API_KEY"
-
-    def _completion_extra_kwargs(self) -> dict:
-        return {
-            "reasoning_effort": "none",
-            **super()._completion_extra_kwargs(),
-        }
-
-
-class GeminiClient(SpanMarkingLLMClient):
-
-    BASE_URL = "http://172.23.205.120:4000/v1"
-    API_KEY_NAME = "LITELLM_API_KEY"
-
-    def _completion_extra_kwargs(self) -> dict:
-        return {
-            "reasoning_effort": "minimal",
-            **super()._completion_extra_kwargs(),
-        }
-
-
-class AzureClient(SpanMarkingLLMClient):
-
-    API_KEY_NAME = "AZURE_API_KEY"
-
-    def __init__(self, *args, **kwargs):
-        self.BASE_URL = os.environ.get("AZURE_BASE_URL")
-        super().__init__(*args, **kwargs)
-
-    def _completion_extra_kwargs(self) -> dict:
-        return {
-            "reasoning_effort": "none",
-            **super()._completion_extra_kwargs(),
-        }
-
-
-class DeepseekClient(SpanMarkingLLMClient):
-
-    BASE_URL = "https://api.deepseek.com"
-    API_KEY_NAME = "DEEPSEEK_API_KEY"
-
-    def _completion_extra_kwargs(self) -> dict:
-        return {
-            "reasoning_effort": "high",
-            "extra_body": {"thinking": {"type": "enabled"}},
-            **super()._completion_extra_kwargs(),
-        }

preamble.md

@@ -1,3 +0,0 @@
-# Generative semantic diff
-
-Demo for **generative semantic difference recognition**: compare auxiliary synchronization (LLM minimal edit, then sequence diff) and span marking (LLM `{{ }}` markers) on the same parallel texts.