# Benchmarks
Real-world compression metrics and performance data.
## Token Compression
Tested with Telegram export (34,478 messages), measured with OpenAI tokenizer (cl100k_base).
### By Output Format
| Raw JSON | 11,177,258 | - | baseline | 1x |
| **CSV** | - | 849,915 | **92.4%** | **13.2x** |
| JSONL | - | 1,029,130 | 90.8% | 10.9x |
| JSON | - | 1,333,586 | 88.1% | 8.4x |
### Why CSV wins
| Brackets `{}[]` | None | `[{},{}]` | `{}` per line |
| Key names | Header only | Every message | Every message |
| Quotes | Minimal | Every string | Every string |
| Delimiter | `;` (1 char) | `,` + spaces | newline |
### Estimated cost savings (GPT-4o)
| 10K messages | $0.56 | $0.04 | $0.52 |
| 100K messages | $5.60 | $0.43 | $5.17 |
| 1M messages | $56.00 | $4.30 | $51.70 |
*Based on $5/1M input tokens*
---
## Criterion Benchmark Results
All benchmarks run with `cargo bench --bench parsing` on release build.
### Parsing Performance
| **Instagram** | 37.9 µs | 377 µs | 3.6 ms | 18.0 ms |
| **Telegram** | 49.2 µs | 487 µs | 7.2 ms | 32.0 ms |
| **Discord** | 60.1 µs | 567 µs | 5.6 ms | 32.1 ms |
| **WhatsApp** | 2.6 ms | 4.4 ms | 22.3 ms | 102.2 ms |
### Throughput (messages/second)
| **Instagram** | 2.6-2.8 M/s |
| **Telegram** | 1.4-2.0 M/s |
| **Discord** | 1.5-1.8 M/s |
| **WhatsApp** | 38-489 K/s |
> WhatsApp uses regex-based text parsing, hence slower than JSON parsers.
### Operations Performance
| **Merge consecutive** | 8.9 µs | 84 µs | 690 µs | 7.0 ms |
| **Filter by sender** | 8.2 µs | 75 µs | 744 µs | 7.4 ms |
| **Filter by date** | 8.1 µs | 77 µs | 764 µs | 7.4 ms |
| Merge consecutive | 11-14 M/s |
| Filter by sender | 12-13 M/s |
| Filter by date | 12-13 M/s |
### Output Format Performance
| **CSV** | 8.1 µs | 77 µs | 874 µs |
| **JSONL** | 10.4 µs | 102 µs | 998 µs |
| **JSON** | 16.6 µs | 158 µs | 1.5 ms |
| CSV | 11-12 M/s |
| JSONL | 9-10 M/s |
| JSON | 6.0-6.6 M/s |
### Full Pipeline (Parse → Merge → Filter → Output)
| 1K | 602 µs | 1.66 M/s |
| 10K | 5.9 ms | 1.70 M/s |
| 50K | 29.8 ms | 1.68 M/s |
---
## Message Merging
Consecutive messages from the same sender are merged into one entry.
### Real-world results
| Telegram group | 34,478 | 26,169 | 24% |
| WhatsApp personal | 1,751 | 809 | 54% |
| Instagram DM | 3,292 | 1,660 | 50% |
| Discord channel | 1,232 | 583 | 53% |
### When merging helps most
- **Group chats with few participants** - people send many short messages
- **Personal chats** - rapid back-and-forth
- **Voice message transcripts** - often split into fragments
- **Discord announcements** - admins often send consecutive updates
---
## Format Comparison
### For LLM context (ChatGPT/Claude)
| Token efficiency | Best | Low | Low |
| Readability | Good | Best | Good |
| Copy-paste friendly | Best | Good | Low |
| **Recommendation** | **Best** | OK | Not ideal |
### For RAG/Vector DB
| One doc per line | No | No | Best |
| Streaming parse | Low | No | Best |
| Schema flexibility | Low | Best | Best |
| **Recommendation** | Not ideal | OK | **Best** |
### For archival/analysis
| Full metadata | Good | Best | Best |
| Excel/Sheets | Best | No | No |
| jq/scripting | Low | Best | Best |
| **Recommendation** | For spreadsheets | **Best** | Also good |
---
## Stress Test Results
Toxic data generator with:
- 100KB+ messages
- Zalgo text (combining characters)
- 50+ emojis per message
- Special characters: `;`, `"`, `\n`, `\t`
- Unicode: Cyrillic, Japanese, Arabic
- Empty/whitespace messages
### Results
| No crashes | Pass |
| CSV escaping correct | Pass |
| Unicode preserved | Pass |
| Empty filtered | Pass |
| Throughput | 17-24K msg/s |
---
## Run Your Own Benchmarks
```bash
# Run all benchmarks
cargo bench --bench parsing
# Run specific benchmark
cargo bench --bench parsing -- telegram_parsing
# Save baseline for comparison
cargo bench --bench parsing -- --save-baseline main
# Compare against baseline
cargo bench --bench parsing -- --baseline main
```
---
*Last updated: December 2025*