ASEED (AI Search Engine Evaluation Dataset) is a dataset rewritten based on real user search scenarios, which is used to evaluate the question-answering quality of current mainstream AI search engines.
We plan to create a standardized and comprehensive dataset to evaluate AI search engines comprehensively and accurately. Considering the language diversity of the global user community, our evaluation report covers multiple languages including English, Japanese, Simplified Chinese, Russian, and Traditional Chinese, and has conducted preliminary testing and evaluation on the accuracy of these question-answering search engine products.
| Product Name | Accuracy | Answer correctness |
|---|---|---|
| Metaso | 66.00% | 62.70% |
| Perplexity | 65.00% | 65.22% |
| iAsk | 60.00% | 62.27% |
| You.com | 34.00% | 58.27% |
The latest report is: evaluation-report-v1.1
You can check the Release to access historical reports.
Scenario Distribution
| Search Scenario | Number of Cases | Percentage of Total |
|---|---|---|
| Technical Consulting | 20 | 20% |
| Real-Time News | 20 | 20% |
| Local Search | 20 | 20% |
| Product Search | 20 | 20% |
| Business Consulting | 20 | 20% |
| Total | 100 | 100% |
Language Distribution
| Evaluation Languages | Language Proportion |
|---|---|
| English | 20% |
| Japanese | 20% |
| Simplified Chinese | 20% |
| Traditional Chinese | 20% |
| Russian | 10% |
| Korean | 10% |
| Total | 100% |
Data Sample
id|query|ground_truth|assertion_type|assertion_value|language|description ---|---|---|---|---|--- 1|What is Toyota's biggest risk in the future?|Toyota's biggest risk in the future is the consumer adoption of battery electric vehicles (BEVs)|llm-rubric|the consumer adoption of battery electric vehicles (BEVs)|en|-| 2|海康威視去年的業績如何|在2023年，海康威視的業績恢復增長，公司實現了營業收入893.55億元，同比增長7.44%，淨利潤為141.17億元，同比增長9.96%。|llm-rubric|在2023年，海康威視實現了營業收入893.**億元|zh_tw|-| 3|最近台灣地震造成的傷亡人數是多少？|截至2024年4月25日下午，該次地震在全台造成18人罹難|llm-rubric|截至2024年4月25日下午，該次地震在全台造成18人罹難|ja|-|
Testing Method
-
Accuracy：Manual Evaluation
The actual answers are manually compared with the Ground truth and Key words. If the answer is a match or relevant, it is marked as "True"; otherwise, it is "False".
-
Answer Correctness: GPT-4 Scoring
The Answer correctness metric from the Ragas platform, with the GPT-4-Turbo model being used for evaluation. The formula for calculating “Answer correctness” is as follows.
(Metric Description： https://docs.ragas.io/en/latest/concepts/metrics/answer_correctness.html)
Evaluation Metrics
| Metric | Definition of Evaluation | Advantages |
|---|---|---|
| Accuracy | Manual evaluation of the actual answers based on Ground truth and key words | Precise assessment |
| Answer correctness | GPT-4 comparison of Ground truth with actual answers | Less labor-intensive |
Product Selection
To conduct this performance evaluation of question-answering search engines, we referred to the list of leading AI search engine products provided by the aicpb.com website. After excluding traditional keyword search engines, we carefully selected four of the most outstanding AI question-answering search engine products on the market for our evaluation.
The specific information about the aforementioned four products is as follows (referenced from statistics at aicpb.com):
| Rank | Product Name | Category | Monthly Visits (March) | MoM Change |
|---|---|---|---|---|
| 2 | Perplexity | AI Search Engine | 64.14M | 25.17% |
| 3 | You.com | AI Search Engine | 10.44M | 14.61% |
| 5 | Metaso | AI Search Engine | 7.21M | 551.36% |
| 7 | iAsk | AI Search Engine | 3.43M | 15.43% |
Detailed test results
- Overall, the performance of the evaluated products did not meet our expectations. However, it is noteworthy that Metaso, a company primarily serving the Chinese market, performed the best overall in the evaluation, slightly surpassing Perplexity.
- The comprehensive data analysis of all products indicates that the highest accuracy is achieved in answering questions in English. By contrast, Russian questions have the lowest accuracy rate, and the accuracy for Japanese questions is also relatively low.
- In terms of performance by language, Perplexity leads by a significant margin in answering English questions, and its performance in Simplified Chinese is also quite impressive. Metaso stands out in its performance in both Simplified and Traditional Chinese. However, both companies did not reach a satisfactory level in other languages. iAsk demonstrated relatively balanced capabilities across different languages, but overall, it falls within the medium range. You.com, on the other hand, performed well only in answering questions in English.
In the future, we plan to expand our evaluation test set, adding more languages and query scenarios to provide a more comprehensive and detailed performance assessment. We believe that with technological advancements and the enrichment of datasets, AI question-answering search engines will be able to better meet the diverse needs of global users and play an even more important role in the future of search technology.
We are very grateful to Ragas for providing the scoring system.