Incremental eval of cognee pipeline (#445)

* QA eval dataset as argument, with hotpot and 2wikimultihop as options. Json schema validation for datasets.

* Load dataset file by filename, outsource utilities

* restructure metric selection

* Add comprehensiveness, diversity and empowerment metrics

* add promptfoo as an option

* refactor RAG solution in eval;2C

* LLM as a judge metrics implemented in a uniform way

* Use requests.get instead of wget

* clean up promptfoo config template

* minor fixes

* get promptfoo path instead of hardcoding

* minor fixes

* Add LLM as a judge prompts

* Support 4 different rag options in eval

* Minor refactor and logger usage

* feat: make tasks a configurable argument in the cognify function

* Run eval on a set of parameters and save results as json and png

* fix: add data points task

* script for running all param combinations

* enable context provider to get tasks as param

* bugfix in simple rag

* Incremental eval of cognee pipeline

* potential fix: single asyncio run

* temp fix: exclude insights

* Remove insights, have single asyncio run, refactor

* minor fixes

* handle pipeline slices in utils

* include all options in params json

---------

Co-authored-by: lxobr <122801072+lxobr@users.noreply.github.com>
Co-authored-by: hajdul88 <52442977+hajdul88@users.noreply.github.com>

evals/eval_on_hotpot.py

@@ -9,7 +9,7 @@ from cognee.infrastructure.llm.get_llm_client import get_llm_client
 from cognee.infrastructure.llm.prompts import read_query_prompt, render_prompt
 from evals.qa_dataset_utils import load_qa_dataset
 from evals.qa_metrics_utils import get_metrics
-from evals.qa_context_provider_utils import qa_context_providers
+from evals.qa_context_provider_utils import qa_context_providers, create_cognee_context_getter
 
 logger = logging.getLogger(__name__)
 
@@ -94,14 +94,29 @@ async def eval_on_QA_dataset(
     return results
 
 
-if __name__ == "__main__":
+async def incremental_eval_on_QA_dataset(
+    dataset_name_or_filename: str, num_samples, metric_name_list
+):
+    pipeline_slice_names = ["base", "extract_chunks", "extract_graph", "summarize"]
+
+    incremental_results = {}
+    for pipeline_slice_name in pipeline_slice_names:
+        results = await eval_on_QA_dataset(
+            dataset_name_or_filename, pipeline_slice_name, num_samples, metric_name_list
+        )
+        incremental_results[pipeline_slice_name] = results
+
+    return incremental_results
+
+
+async def main():
     parser = argparse.ArgumentParser()
 
     parser.add_argument("--dataset", type=str, required=True, help="Which dataset to evaluate on")
     parser.add_argument(
         "--rag_option",
         type=str,
-        choices=qa_context_providers.keys(),
+        choices=list(qa_context_providers.keys()) + ["cognee_incremental"],
         required=True,
         help="RAG option to use for providing context",
     )
@@ -110,7 +125,18 @@ if __name__ == "__main__":
 
     args = parser.parse_args()
 
-    avg_scores = asyncio.run(
-        eval_on_QA_dataset(args.dataset, args.rag_option, args.num_samples, args.metrics)
-    )
+    if args.rag_option == "cognee_incremental":
+        avg_scores = await incremental_eval_on_QA_dataset(
+            args.dataset, args.num_samples, args.metrics
+        )
+
+    else:
+        avg_scores = await eval_on_QA_dataset(
+            args.dataset, args.rag_option, args.num_samples, args.metrics
+        )
+
     logger.info(f"{avg_scores}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

evals/official_hotpot_metrics.py

@@ -4,11 +4,8 @@ These are the official evaluation metrics for HotpotQA taken from https://hotpot
 
 import re
 import string
-import sys
 from collections import Counter
 
-import ujson as json
-
 
 def normalize_answer(s):
     def remove_articles(text):

evals/qa_context_provider_utils.py

@@ -3,35 +3,49 @@ from cognee.api.v1.search import SearchType
 from cognee.infrastructure.databases.vector import get_vector_engine
 from cognee.modules.retrieval.brute_force_triplet_search import brute_force_triplet_search
 from cognee.tasks.completion.graph_query_completion import retrieved_edges_to_string
+from functools import partial
+from cognee.api.v1.cognify.cognify_v2 import get_default_tasks
 
 
 async def get_raw_context(instance: dict) -> str:
     return instance["context"]
 
 
-async def cognify_instance(instance: dict):
+async def cognify_instance(instance: dict, task_indices: list[int] = None):
     await cognee.prune.prune_data()
     await cognee.prune.prune_system(metadata=True)
-
     for title, sentences in instance["context"]:
         await cognee.add("\n".join(sentences), dataset_name="QA")
-    await cognee.cognify("QA")
+    all_cognify_tasks = await get_default_tasks()
+    if task_indices:
+        selected_tasks = [all_cognify_tasks[ind] for ind in task_indices]
+    else:
+        selected_tasks = all_cognify_tasks
+    await cognee.cognify("QA", tasks=selected_tasks)
 
 
-async def get_context_with_cognee(instance: dict) -> str:
-    await cognify_instance(instance)
+async def get_context_with_cognee(
+    instance: dict,
+    task_indices: list[int] = None,
+    search_types: list[SearchType] = [SearchType.SUMMARIES, SearchType.CHUNKS],
+) -> str:
+    await cognify_instance(instance, task_indices)
 
-    # TODO: Fix insights
-    # insights = await cognee.search(SearchType.INSIGHTS, query_text=instance["question"])
-    summaries = await cognee.search(SearchType.SUMMARIES, query_text=instance["question"])
-    # search_results = insights + summaries
-    search_results = summaries
+    search_results = []
+    for search_type in search_types:
+        search_results += await cognee.search(search_type, query_text=instance["question"])
 
     search_results_str = "\n".join([context_item["text"] for context_item in search_results])
 
     return search_results_str
 
 
+def create_cognee_context_getter(
+    task_indices=None, search_types=[SearchType.SUMMARIES, SearchType.CHUNKS]
+):
+    return partial(get_context_with_cognee, task_indices=task_indices, search_types=search_types)
+
+
 async def get_context_with_simple_rag(instance: dict) -> str:
     await cognee.prune.prune_data()
     await cognee.prune.prune_system(metadata=True)
@@ -57,9 +71,19 @@ async def get_context_with_brute_force_triplet_search(instance: dict) -> str:
     return search_results_str
 
 
+valid_pipeline_slices = {
+    "base": [0, 1, 5],
+    "extract_chunks": [0, 1, 2, 5],
+    "extract_graph": [0, 1, 2, 3, 5],
+    "summarize": [0, 1, 2, 3, 4, 5],
+}
+
 qa_context_providers = {
     "no_rag": get_raw_context,
     "cognee": get_context_with_cognee,
     "simple_rag": get_context_with_simple_rag,
     "brute_force": get_context_with_brute_force_triplet_search,
+} | {
+    name: create_cognee_context_getter(task_indices=slice)
+    for name, slice in valid_pipeline_slices.items()
 }

evals/qa_eval_parameters.json

@@ -3,8 +3,9 @@
         "hotpotqa"
     ],
     "rag_option": [
-        "no_rag",
+        "cognee_incremental",
         "cognee",
+        "no_rag",
         "simple_rag",
         "brute_force"
     ],

evals/qa_eval_utils.py

@@ -44,7 +44,7 @@ def save_results_as_image(results, out_path):
         for num_samples, table_data in num_samples_data.items():
             df = pd.DataFrame.from_dict(table_data, orient="index")
             df.index.name = f"Dataset: {dataset}, Num Samples: {num_samples}"
-            image_path = Path(out_path) / Path(f"table_{dataset}_{num_samples}.png")
+            image_path = out_path / Path(f"table_{dataset}_{num_samples}.png")
             save_table_as_image(df, image_path)