cognee/evals

QA Evaluation

Repeated runs of QA evaluation on 24-item HotpotQA subset, comparing Mem0, Graphiti, LightRAG, and Cognee (multiple retriever configs). Uses Modal for distributed benchmark execution.

Dataset

• hotpot_qa_24_corpus.json and hotpot_qa_24_qa_pairs.json
• hotpot_qa_24_instance_filter.json for instance filtering

Systems Evaluated

• Mem0: OpenAI-based memory QA system
• Graphiti: LangChain + Neo4j knowledge graph QA
• LightRAG: Falkor's GraphRAG-SDK
• Cognee: Multiple retriever configurations (GRAPH_COMPLETION, GRAPH_COMPLETION_COT, GRAPH_COMPLETION_CONTEXT_EXTENSION)

Project Structure

• src/ - Analysis scripts and QA implementations
• src/modal_apps/ - Modal deployment configurations
• src/qa/ - QA benchmark classes
• src/helpers/ and src/analysis/ - Utilities

Notes:

• Use PyProject.toml for dependencies
• Ensure Modal CLI is configured
• Modular QA benchmark classes enable parallel execution on other platforms beyond Modal

Running Benchmarks (Modal)

Execute repeated runs via Modal apps:

• modal run modal_apps/modal_qa_benchmark_<system>.py

Where <system> is one of: mem0, graphiti, lightrag, cognee

Raw results stored in Modal volumes under /qa-benchmarks/<benchmark>/{answers,evaluated}

Results Analysis

• python run_cross_benchmark_analysis.py
• Downloads Modal volumes, processes evaluated JSONs
• Generates per-benchmark CSVs and cross-benchmark summary
• Use visualize_benchmarks.py to create comparison charts

Results

• 45 evaluation cycles on 24 HotPotQA questions with multiple metrics (EM, F1, DeepEval Correctness, Human-like Correctness)
• Significant variance observed in metrics across small runs due to LLM-as-judge inconsistencies
• Cognee showed consistent improvements across all measured dimensions compared to Mem0, Lightrag, and Graphiti

Visualization Results

The following charts visualize the benchmark results and performance comparisons:

Comprehensive Metrics Comparison

A comprehensive comparison of all evaluated systems across multiple metrics, showing Cognee's performance relative to Mem0, Graphiti, and LightRAG.

Optimized Cognee Configurations

Performance analysis of different Cognee retriever configurations (GRAPH_COMPLETION, GRAPH_COMPLETION_COT, GRAPH_COMPLETION_CONTEXT_EXTENSION), showing optimization results.

Notes

• Traditional QA metrics (EM/F1) miss core value of AI memory systems - measure letter/word differences rather than information content
• HotPotQA benchmark mismatch - designed for multi-hop reasoning but operates in constrained contexts vs. real-world cross-context linking
• DeepEval variance - LLM-as-judge evaluation carries inconsistencies of underlying language model