[Eval] Hybrid search evaluation

[JasonLo]

After issue #76 (elastic search integration), we should examine how it affects the evaluation.

Test set: ta1

Comparisons:

1. XDD articles V2
2. Retriever V0
3. Retriever Hybrid (new)
4. Retriever Hybrid + terms pre-filtering

Takeaways:

• In most cases, the Hybrid performs well, yielding identical or similar results in (3) and (4).
• In 10-20% of cases, terms pre-filtering is too stringent and returns no results.
• We still have challenges with key terms, such as:
  • SV2AIR3 model formula
  • What is the SIDARTHE-V model?
  • Differences between the original SIDARTHE and SIDARTHE-V
• Further work is needed on evaluation questions without key terms. Current performance is neither better nor worse than XDD V2. Improved metrics are needed to quantify results.

Source: details