Switched to Claude 3.5 Haiku for both agent and job call summary, seemes to be the best so far

.github/workflows/python_ci.yml

@@ -25,7 +25,7 @@ jobs:
     - name: Test with unittest
       env:
         OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-        MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
+        ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
       run: |
         python -m unittest tests/test_gradio.py
         python -m unittest tests/test_linkedin_resume.py

configuration.py

@@ -6,30 +6,30 @@ from smolagents import OpenAIServerModel
 
 DEFAULT_GITHUB_PROFILE = "https://github.com/gperdrizet"
 
-# AGENT_MODEL = OpenAIServerModel(
-#     model_id="gpt-4.1",
-#     max_tokens=8000
-# )
-
 # Will be used for single shot summarization with no-frills prompting
 # (e.g. job call extraction). It needs to output JSON formatted text,
 # but this task does not require any complex reasoning or planning.
-SUMMARIZER_MODEL = OpenAI(
-    base_url="https://gperdrizet--llama-3-1-8b-instruct-serve.modal.run/v1",
-    api_key=os.environ['MODAL_TOKEN_SECRET']
+SUMMARIZER_CLIENT = OpenAI(
+    base_url="https://api.anthropic.com/v1/",
+    api_key=os.environ["ANTHROPIC_API_KEY"]
 )
 
+SUMMARIZER_MODEL = "claude-3-5-haiku-20241022"
+
 # Will be used for resume resume writing agent via HuggingFace smolagents
 # Including selection of relevant projects from GitHub profile
 #
 # Notes:
 # - DeepSeek-R1-Distill-Qwen-32B does not seem to work well with smolagents,
 #   has trouble correctly formatting responses as code.
+# - Qwen2.5-Coder-14B-Instruct works OK, but is not great at markdown formatting
+#   and tends to get some details wrong.
+# - Claude-3-5-Haiku is the best model for this task so far.
 
 AGENT_MODEL = OpenAIServerModel(
-    model_id="Qwen/Qwen2.5-Coder-14B-Instruct", # Same as HF model string
-    api_base="https://gperdrizet--qwen2-5-coder-14b-instruct-serve.modal.run/v1",
-    api_key=os.environ["MODAL_TOKEN_SECRET"],
+    model_id="claude-3-5-haiku-20241022", # Same as HF model string
+    api_base="https://api.anthropic.com/v1/",
+    api_key=os.environ["ANTHROPIC_API_KEY"],
 )
 
 INSTRUCTIONS = """

functions/job_call.py

@@ -4,7 +4,11 @@ import json
 import logging
 from pathlib import Path
 from datetime import datetime
-from configuration import JOB_CALL_EXTRACTION_PROMPT, SUMMARIZER_MODEL
+from configuration import (
+    JOB_CALL_EXTRACTION_PROMPT,
+    SUMMARIZER_MODEL,
+    SUMMARIZER_CLIENT
+)
 
 # pylint: disable=broad-exception-caught
 
@@ -57,11 +61,6 @@ def summarize_job_call(job_call: str) -> str:
 
     logger.info("Summarizing job call (%d characters)", len(job_call))
 
-    # Default to first available model
-    model = SUMMARIZER_MODEL.models.list().data[0]
-    model_id = model.id
-    print(f"Using model: {model_id}")
-
     messages = [
         {
             'role': 'system',
@@ -70,12 +69,12 @@ def summarize_job_call(job_call: str) -> str:
     ]
 
     completion_args = {
-        'model': model_id,
+        'model': SUMMARIZER_MODEL,
         'messages': messages,
     }
 
     try:
-        response = SUMMARIZER_MODEL.chat.completions.create(**completion_args)
+        response = SUMMARIZER_CLIENT.chat.completions.create(**completion_args)
 
     except Exception as e:
         response = None