reporeformat V2 (#23)

* reformat factor implement process * move some code to more reasonable place * fix the bug * add test function in factor_extract_and_implement.py * change select factor number to ratio , add some factor implement setting and fix some bug while using knowledgebase * change evoagent * add abstract class EvoAgent * add benchmark workflow * fix some bug in llm_utils * run wenjun's code * fix the knowledgebase instance check --------- Co-authored-by: xuyang1 <xuyang1@microsoft.com>
2026-08-01 01:27:42 +00:00 · 2024-06-14 12:59:44 +08:00
parent 9e82da243b
commit ebb659a018
32 changed files with 2227 additions and 1141 deletions
@@ -0,0 +1,69 @@
+from rdagent.oai.llm_utils import APIBackend
+from jinja2 import Template
+from rdagent.factor_implementation.share_modules.factor_implementation_config import FactorImplementSettings
+import json
+from rdagent.factor_implementation.share_modules.factor_implementation_utils import get_data_folder_intro
+from rdagent.factor_implementation.evolving.factor import FactorEvovlingItem
+from rdagent.core.prompts import Prompts
+from pathlib import Path
+
+scheduler_prompts = Prompts(file_path=Path(__file__).parent.parent / "prompts.yaml")
+
+def RandomSelect(to_be_finished_task_index, implementation_factors_per_round):
+    import random
+    to_be_finished_task_index = random.sample(
+        to_be_finished_task_index,
+        implementation_factors_per_round,
+    )
+    print("The random selection is:",to_be_finished_task_index)
+    return to_be_finished_task_index
+
+def LLMSelect(to_be_finished_task_index, implementation_factors_per_round, evo:FactorEvovlingItem, former_trace):
+    tasks = []
+    for i in to_be_finished_task_index:
+    # find corresponding former trace for each task
+        target_factor_task_information = evo.target_factor_tasks[i].get_factor_information()
+        if target_factor_task_information in former_trace:
+            tasks.append((i, evo.target_factor_tasks[i], former_trace[target_factor_task_information]))
+
+    system_prompt = Template(
+        scheduler_prompts["select_implementable_factor_system"],
+    ).render(
+        data_info=get_data_folder_intro(),
+    )
+
+    session = APIBackend(use_chat_cache=False).build_chat_session(
+        session_system_prompt=system_prompt,
+    )
+
+    while True:
+        user_prompt = (
+            Template(
+                scheduler_prompts["select_implementable_factor_user"],
+            )
+            .render(
+                factor_num = implementation_factors_per_round,
+                target_factor_tasks=tasks,
+            )
+        )
+        if (
+            session.build_chat_completion_message_and_calculate_token(
+                user_prompt,
+            )
+            < FactorImplementSettings().chat_token_limit
+        ):
+            break
+
+    response = session.build_chat_completion(
+        user_prompt=user_prompt,
+        json_mode=True,
+    )
+    try:
+        selection = json.loads(response)["selected_factor"]
+        if not isinstance(selection, list):
+            return to_be_finished_task_index
+        selection_index = [x for x in selection if isinstance(x, int)]
+    except:
+        return to_be_finished_task_index
+    
+    return selection_index