BENCHMARK · MESA 0.9.3
Turkish legal retrieval benchmark
Same retrieval quality, faster ingest: with the built-in reranker, 0.9.2 and 0.9.3 both reach Hit@5 74.4% and MRR 0.656 at the displayed precision.
74.4%
Hit@5
Correct passage in the first five results
57.9%
Hit@1
Correct passage in the first result
0.656
MRR
Ranking of the first correct passage
100.0%
Citation Hit@5
Explicit citation questions
MESA 0.9.3 · BGE-M3 + built-in BGE reranker (T2)
Compare results
Models off
BGE-M3 embeddings
Built-in reranker
| Configuration | 0.9.2 | 0.9.3 | Hit@5 change | ||
|---|---|---|---|---|---|
| Hit@5 | MRR | Hit@5 | MRR | ||
| Models off | 39.7% | 0.300 | 38.0% | 0.297 | -1.7 pp |
| BGE-M3 embeddings | 67.8% | 0.514 | 67.8% | 0.522 | 0.0 pp |
| Built-in reranker | 74.4% | 0.656 | 74.4% | 0.656 | 0.0 pp |
Changes are in percentage points, calculated as selected version minus reference. Profiles are matched by model setup. Search settings and hardware can differ. See the individual reports for timing and offline experiments.
Hit@5 by configuration
Search results
November 18, 2026 · 141 documents · 121 / 145 answerable test questions
| Configuration | Hit@5 | Hit@1 | MRR | Hit@10 | Doc@1 | Doc@5 | p50 | p95 | Hardware |
|---|---|---|---|---|---|---|---|---|---|
| 0.9.0 · Keyword search (Mode A)20261029T190820Z | 33.1% | 22.3% | 0.293 | 38.8% | 36.4% | 60.3% | 0.96 s | 1.93 s | Colab T4 |
| 0.9.0 · BGE-M3, equal weights (S0)20261029T234645Z_exp | 51.2% | 34.7% | 0.431 | 60.3% | 58.7% | 70.3% | 1.91 s | 3.49 s | Colab A100 |
| 0.9.0 · BGE-M3, tuned vector weight (S4)20261029T234645Z_exp | 62.0% | 39.7% | 0.500 | 70.3% | 60.3% | 75.2% | 1.87 s | 3.29 s | Colab A100 |
| 0.9.2 · Models off (T0)20261107T141454Z | 39.7% | 22.3% | 0.300 | 46.3% | 44.6% | 57.9% | 0.45 s | 0.94 s | Colab A100 |
| 0.9.2 · Magibu embeddings (T1m)20261107T141454Z | 50.4% | 32.2% | 0.410 | 59.5% | 52.9% | 68.6% | 1.39 s | 2.33 s | Colab A100 |
| 0.9.2 · BGE-M3 embeddings (T1b)20261107T141454Z | 67.8% | 38.0% | 0.514 | 76.9% | 56.2% | 76.9% | 1.35 s | 2.27 s | Colab A100 |
| 0.9.2 · BGE-M3 + built-in BGE reranker (T2)20261107T141454Z | 74.4% | 57.9% | 0.656 | 81.0% | 69.4% | 79.3% | 1.98 s | 2.98 s | Colab A100 |
| 0.9.3 · Models off (T0)20261117T230132Z_v093 | 38.0% | 22.3% | 0.297 | 47.9% | 43.8% | 57.9% | 0.45 s | 0.94 s | Colab A100-SXM4-80GB |
| 0.9.3 · BGE-M3 embeddings (T1b)20261117T230132Z_v093 | 67.8% | 38.8% | 0.522 | 77.7% | 57.9% | 77.7% | 1.37 s | 2.36 s | Colab A100-SXM4-80GB |
| 0.9.3 · BGE-M3 + built-in BGE reranker (T2)20261117T230132Z_v093 | 74.4% | 57.9% | 0.656 | 81.0% | 69.4% | 79.3% | 2.02 s | 3.00 s | Colab A100-SXM4-80GB |
| Configuration | Hit@5 | Hit@1 | MRR | Hit@10 | Doc@1 | Doc@5 | p50 | p95 | Hardware |
|---|---|---|---|---|---|---|---|---|---|
| 0.9.0 · Article chunks + offline Jina20261029T234645Z_exp | 76.0% | 57.0% | 0.654 | 80.2% | 69.4% | 82.6% | — | — | Colab A100 |
- 0.9.0: Mode A ran on a T4. The S0 and S4 rows use the A100 reruns from the article-chunking study. The original T4 embedding runs are separate measurements.
- 0.9.0: The Jina reranking ran offline over saved results. It is reported separately from built-in configurations and excluded from the comparison.
- 0.9.2: All rows ran on a Colab A100. T2 reused the T1b store, so it has no separate ingestion measurement.
- 0.9.3: All rows ran on a Colab A100-SXM4-80GB. T2 reused the T1b store, so it has no separate ingestion measurement.
Results by question type
Hit@5 per question type. Darker cells mean the right passage was found more often.
| Question type | 0.9.0 · Keyword search (Mode A) | 0.9.0 · BGE-M3, equal weights (S0) | 0.9.0 · BGE-M3, tuned vector weight (S4) | 0.9.0 · Article chunks + offline Jina | 0.9.2 · Models off (T0) | 0.9.2 · Magibu embeddings (T1m) | 0.9.2 · BGE-M3 embeddings (T1b) | 0.9.2 · BGE-M3 + built-in BGE reranker (T2) | 0.9.3 · Models off (T0) | 0.9.3 · BGE-M3 embeddings (T1b) | 0.9.3 · BGE-M3 + built-in BGE reranker (T2) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Exact wording (36) | 66.7% | 86.1% | 86.1% | 94.4% | 75.0% | 77.8% | 86.1% | 97.2% | 69.4% | 88.9% | 97.2% |
| Paraphrase (39) | 10.3% | 28.2% | 56.4% | 79.5% | 15.4% | 48.7% | 69.2% | 71.8% | 18.0% | 69.2% | 71.8% |
| Cross-document (17) | 11.8% | 47.1% | 47.1% | 70.6% | 35.3% | 23.5% | 58.8% | 76.5% | 35.3% | 52.9% | 76.5% |
| Number, date or docket (29) | 34.5% | 41.4% | 48.3% | 51.7% | 31.0% | 34.5% | 48.3% | 48.3% | 27.6% | 48.3% | 48.3% |
Citation questions
0.9.3 keeps the 0.9.2 result: explicit citation Hit@5 is 100.0% (50/50) for T0, T1b and T2. Natural wording is 97.5% (39/40) for T0 and T1b, and 100.0% (40/40) for T2.
| Configuration | Explicit citation (50) | Natural wording (40) |
|---|---|---|
| 0.9.2 · Models off (T0) | 100.0% | 97.5% |
| 0.9.2 · Magibu embeddings (T1m) | 100.0% | 95.0% |
| 0.9.2 · BGE-M3 embeddings (T1b) | 100.0% | 97.5% |
| 0.9.2 · BGE-M3 + built-in BGE reranker (T2) | 100.0% | 100.0% |
| 0.9.3 · Models off (T0) | 100.0% | 97.5% |
| 0.9.3 · BGE-M3 embeddings (T1b) | 100.0% | 97.5% |
| 0.9.3 · BGE-M3 + built-in BGE reranker (T2) | 100.0% | 100.0% |
Time until searchable
| Configuration | Total time | Chunks | Seconds per chunk | Hardware |
|---|---|---|---|---|
| 0.9.0 · BGE-M3, equal weights (S0) | 29.7 min | 3,480 | 0.51 s | Colab A100 |
| 0.9.2 · Models off (T0) | 15.8 min | 5,041 | 0.19 s | Colab A100 |
| 0.9.2 · Magibu embeddings (T1m) | 23.7 min | 5,041 | 0.28 s | Colab A100 |
| 0.9.2 · BGE-M3 embeddings (T1b) | 20.7 min | 5,041 | 0.25 s | Colab A100 |
| 0.9.3 · Models off (T0) | 7.4 min | 5,041 | 0.09 s | Colab A100-SXM4-80GB |
| 0.9.3 · BGE-M3 embeddings (T1b) | 12.0 min | 5,041 | 0.14 s | Colab A100-SXM4-80GB |
0.9.3 batches ingest work; retrieval quality is unchanged.
Download results
MESA 0.9.0 · Report and aggregate metrics (JSON) ↓
How we measured
Tier M contains 141 Turkish laws, regulations and court decisions. A language model generated questions from the corpus. A separate model and rule checks reviewed them. Configurations were chosen on 101 development questions before evaluation on 145 held-out questions. Retrieval scores use the 121 answerable questions.
What do the metrics mean?
- Hit@1 · Hit@5 · Hit@10
- Share of questions with the correct passage in the first 1, 5 or 10 results.
- MRR
- Average reciprocal rank of the first correct passage. A score of 1 means it always came first.
- Doc@1 · Doc@5
- Share of questions with the correct document among the first 1 or 5 results.
- p50 · p95
- Half of queries finished within p50 and 95% within p95. Both are measured in seconds.
Scope and limitations
The test measures whether MESA retrieves the correct passage or document. The 24 questions with no answer in the corpus are outside the retrieval scores. The corpus is small and its decisions come from the Court of Cassation. The questions are synthetic.
For number, date and docket questions, the T2 result is unchanged in 0.9.3: 14/29 = 48.3% Hit@5. The benchmark requests 50 results. With a 10-result request the reranker sees 20 candidates; on the dev split Hit@5 was 75.3% instead of 77.7%. This comparison changes both the candidate pool and the number of returned results.
0.9.3 run: 18 November 2026 · mesa-memory 0.9.3 · 20261117T230132Z_v093.
