Support OPENAI_BASE_URL routing and JSON parsing

compass/llm/config.py

@@ -200,6 +200,14 @@ def client_kwargs(self):
             for key, env_var in arg_env_pairs:
                 if self._client_kwargs.get(key) is None:
                     self._client_kwargs[key] = os.environ.get(env_var)
+        elif self.client_type == "openai":
+            arg_env_pairs = [
+                ("api_key", "OPENAI_API_KEY"),
+                ("base_url", "OPENAI_BASE_URL"),
+            ]
+            for key, env_var in arg_env_pairs:
+                if self._client_kwargs.get(key) is None:
+                    self._client_kwargs[key] = os.environ.get(env_var)
 
         return self._client_kwargs
 

compass/utilities/parsing.py

@@ -39,23 +39,29 @@ def llm_response_as_json(content):
 
     Returns
     -------
-    dict
+    object
         Parsed JSON structure. When parsing fails, the function returns
         an empty dictionary.
 
     Notes
     -----
     The parser strips Markdown code fences, coerces Python-style
-    booleans to lowercase JSON literals, and logs the raw response on
-    decode failure. The logging includes guidance for increasing token
-    limits or updating prompts.
+    booleans to lowercase JSON literals, and first attempts strict JSON
+    decoding. If strict decoding fails, the parser attempts to recover
+    the first valid JSON object or array embedded in the response. If
+    recovery also fails, the raw response is logged with guidance for
+    prompt/token adjustments.
     """
     content = clean_backticks_from_llm_response(content)
     content = content.removeprefix("json").lstrip("\n")
     content = content.replace("True", "true").replace("False", "false")
     try:
         content = json.loads(content)
     except json.decoder.JSONDecodeError:
+        parsed_content = _parse_first_json_payload(content)
+        if isinstance(parsed_content, dict):
+            return parsed_content
+
         logger.exception(
             "LLM returned improperly formatted JSON. "
             "This is likely due to the completion running out of tokens. "
@@ -68,6 +74,22 @@ def llm_response_as_json(content):
     return content
 
 
+def _parse_first_json_payload(content):
+    """Parse first valid JSON payload embedded in text"""
+    decoder = json.JSONDecoder()
+    for start_ind, start_char in enumerate(content):
+        if start_char not in {"{", "["}:
+            continue
+        try:
+            parsed_content, __ = decoder.raw_decode(content[start_ind:])
+        except json.decoder.JSONDecodeError:
+            continue
+        else:
+            return parsed_content
+
+    return None
+
+
 def merge_overlapping_texts(text_chunks, n=300):
     """Merge text chunks while trimming overlapping boundaries
 

tests/python/unit/llm/test_config.py

@@ -0,0 +1,54 @@
+"""Tests for LLM configuration helpers"""
+
+from pathlib import Path
+
+import pytest
+
+from compass.llm.config import OpenAIConfig
+
+
+def test_openai_client_kwargs_loaded_from_env(monkeypatch):
+    """OpenAI kwargs can be populated from OPENAI_* env vars"""
+    monkeypatch.setenv("OPENAI_API_KEY", "test-openai-key")
+    monkeypatch.setenv("OPENAI_BASE_URL", "https://litellm.example.gov")
+
+    config = OpenAIConfig(name="gpt-4o-mini", client_type="openai")
+
+    assert config.client_kwargs["api_key"] == "test-openai-key"
+    assert config.client_kwargs["base_url"] == "https://litellm.example.gov"
+
+
+def test_openai_client_kwargs_user_values_take_precedence(monkeypatch):
+    """Explicit client kwargs should not be replaced by env vars"""
+    monkeypatch.setenv("OPENAI_API_KEY", "env-key")
+    monkeypatch.setenv("OPENAI_BASE_URL", "https://env.example")
+
+    config = OpenAIConfig(
+        name="gpt-4o-mini",
+        client_type="openai",
+        client_kwargs={
+            "api_key": "user-key",
+            "base_url": "https://user.example",
+        },
+    )
+
+    assert config.client_kwargs["api_key"] == "user-key"
+    assert config.client_kwargs["base_url"] == "https://user.example"
+
+
+def test_azure_client_kwargs_unchanged(monkeypatch):
+    """Azure env var mapping remains unchanged"""
+    monkeypatch.setenv("AZURE_OPENAI_API_KEY", "azure-key")
+    monkeypatch.setenv("AZURE_OPENAI_VERSION", "2024-02-15-preview")
+    monkeypatch.setenv("AZURE_OPENAI_ENDPOINT", "https://azure.example")
+
+    config = OpenAIConfig(name="gpt-4o-mini", client_type="azure")
+
+    assert config.client_kwargs["api_key"] == "azure-key"
+    assert config.client_kwargs["api_version"] == "2024-02-15-preview"
+    assert config.client_kwargs["azure_endpoint"] == "https://azure.example"
+    assert "base_url" not in config.client_kwargs
+
+
+if __name__ == "__main__":
+    pytest.main(["-q", "--show-capture=all", Path(__file__), "-rapP"])

tests/python/unit/utilities/test_utilities_parsing.py

@@ -43,6 +43,14 @@ def test_clean_backticks_from_llm_response(in_str, expected):
         ('{"a": True', {}),
         ('json\n{"key": "value"}', {"key": "value"}),
         ('{"a": True, "b": False}', {"a": True, "b": False}),
+        (
+            (
+                "I can extract date information from the URL provided. "
+                "However, the URL does not contain date information.\n\n"
+                '{"year": null, "month": null, "day": null}'
+            ),
+            {"year": None, "month": None, "day": None},
+        ),
     ],
 )
 def test_llm_response_as_json(in_str, expected):