Metadata-Version: 2.4
Name: less-tokens
Version: 0.2.0
Summary: Cut LLM prompt token costs by 30-40% with deterministic, training-free lexical compression. Shrink prompts for OpenAI, Anthropic, and any LLM API while preserving output quality. Includes zone-aware compression that protects JSON schemas and output formats, async support, and a built-in 6-metric quality evaluator.
Author-email: Shamin Chokshi <shaminchokshi2000@gmail.com>
License: MIT
Project-URL: Homepage, https://github.com/shaminchokshi/less-tokens
Project-URL: Documentation, https://github.com/shaminchokshi/less-tokens#readme
Project-URL: Issues, https://github.com/shaminchokshi/less-tokens/issues
Project-URL: Repository, https://github.com/shaminchokshi/less-tokens
Project-URL: Changelog, https://github.com/shaminchokshi/less-tokens/releases
Keywords: llm,large-language-models,prompt,prompt-compression,prompt-engineering,prompt-optimization,compression,token-compression,token-reduction,token-optimization,tokens,tokenization,reduce-tokens,save-tokens,cost-reduction,cost-optimization,llm-cost,api-cost,openai,gpt,gpt-4,gpt-4o,anthropic,claude,gemini,chatgpt,nlp,natural-language-processing,text-compression,text-processing,stopwords,lemmatization,nltk,wordnet,tiktoken,inference,inference-optimization,latency,efficiency,shortening,summarization,preprocessing,rag,context-compression,bertscore,deterministic,lightweight
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: Intended Audience :: Information Technology
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.9
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Text Processing :: Linguistic
Classifier: Topic :: Text Processing :: General
Classifier: Topic :: Software Development :: Libraries :: Python Modules
Classifier: Typing :: Typed
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: nltk>=3.8
Requires-Dist: tiktoken>=0.6.0
Requires-Dist: rouge-score>=0.1.2
Requires-Dist: sentence-transformers>=2.5.0
Requires-Dist: bert-score>=0.3.13
Requires-Dist: numpy>=1.21
Provides-Extra: dev
Requires-Dist: pytest>=7; extra == "dev"
Requires-Dist: build; extra == "dev"
Requires-Dist: twine; extra == "dev"
Dynamic: license-file

<p align="center">
  <img src="https://raw.githubusercontent.com/shaminchokshi/less-tokens/main/logo.jpeg" alt="less-tokens logo" width="320">
</p>

<h1 align="center">less-tokens</h1>

<p align="center">
  <a href="https://pypi.org/project/less-tokens/"><img src="https://img.shields.io/pypi/v/less-tokens.svg" alt="PyPI version"></a>
  <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-blue.svg" alt="License: MIT"></a>
  <a href="https://www.python.org/downloads/"><img src="https://img.shields.io/badge/python-3.9+-blue.svg" alt="Python 3.9+"></a>
</p>

<p align="center">
  <b>Shrink your LLM prompts by 30 to 40 percent without changing what the model says back.</b>
</p>

`less-tokens` is a small Python library that compresses prompts before you send them to an LLM. It strips out filler words, redundant phrases, and grammatical scaffolding that the model doesn't actually need. The result is a shorter prompt that costs less and responds faster, while producing essentially the same answer.

No neural model, no GPU, no API key for the compression itself. It's classical lexical NLP, runs in milliseconds on a laptop CPU, and is fully deterministic.

```python
from less_tokens import compress

original = "I was wondering if you could please explain to me how I can run a Python script from the command line."
compressed = compress(original,
                      remove_filler_phrases=1,
                      remove_stopwords=1,
                      apply_contractions=1)

print(compressed)
# "explain run Python script command line"
```

## Contents

- [Why this exists](#why-this-exists)
- [Install](#install)
- [The functions at a glance](#the-functions-at-a-glance)
- [compress: shrink a prompt](#compress-shrink-a-prompt)
- [compress_structured: protect the parts that matter](#compress_structured-protect-the-parts-that-matter)
- [compare: measure the quality tradeoff](#compare-measure-the-quality-tradeoff)
- [Async support](#async-support)
- [A complete example](#a-complete-example)
- [Under the hood](#under-the-hood)
- [Limitations](#limitations)

## Why this exists

If you're calling OpenAI, Anthropic, or any other LLM API at meaningful volume, every token has a cost. And typical prompts carry a lot of fat:

- *"I was wondering if you could..."* is hedging the model ignores
- *"the"*, *"a"*, *"is"* are function words that rarely change meaning
- *"basically"*, *"actually"*, *"really"* are fillers
- *"for example"* is just a verbose way to write *"e.g."*

Strip these out and the model still gets your point, but you pay less. On a large benchmark we ran (1,242 prompts, 18,630 paired LLM completions), here's how the headline numbers came out:

| Setting | Token reduction | Output similarity (BERTScore F1) |
|---------|-----------------|----------------------------------|
| Conservative | about 2% | 0.96 |
| **Balanced** | **about 30%** | **0.91** |
| Aggressive | about 35% | 0.91 |
| Maximum | about 40% | 0.88 |

The balanced setting is the sweet spot for most production use. Aggressive gets you a bit more compression without much extra quality loss.

## Install

```bash
pip install less-tokens
```

On first use it downloads about 30 MB of NLTK data automatically. If you also call `compare()`, BERTScore will download an additional ~1 GB model the first time. You can skip that with `bertscore=False` if you don't need it.

Using a virtual environment is highly recommended:

```bash
python -m venv .venv

# Windows
.venv\Scripts\Activate.ps1

# macOS or Linux
source .venv/bin/activate

pip install less-tokens
```

## The functions at a glance

The library gives you four functions. Two for compressing, one for measuring, and async versions for scale.

| Function | What it's for |
|----------|---------------|
| `compress()` | Compress a plain prompt using any combination of techniques |
| `compress_structured()` | Compress a prompt that has parts you must protect, like a JSON output format or strict rules |
| `compare()` | Measure how similar the LLM's two answers are, across six metrics |
| `acompress()` / `acompress_structured()` | Async versions of the two compressors, for use inside an event loop |

If your prompt is just instructions, use `compress()`. If your prompt mixes instructions with an output schema or rules that can't be touched, use `compress_structured()`. Start there.

## compress: shrink a prompt

Pass your prompt and any combination of eleven flags. Each flag is `1` to enable or `0` to disable. Bool and string aliases like `True` or `"on"` work too. Defaults are off for everything except whitespace cleanup, so you choose what runs.

```python
from less_tokens import compress

short = compress(
    "I was wondering if you could explain this to me.",
    remove_filler_phrases=1,
    remove_stopwords=1,
)
# "explain"
```

### The eleven techniques

| Flag | What it does | Example |
|------|--------------|---------|
| `remove_filler_phrases` | Strips hedging phrases | "I was wondering if you could explain" becomes "explain" |
| `apply_abbreviations` | Replaces verbose forms | "for example" becomes "e.g." |
| `apply_contractions` | Combines into contractions | "do not" becomes "don't" |
| `remove_filler_words` | Drops single-word fillers | "this is basically really good" becomes "this is good" |
| `remove_stopwords` | Drops common stopwords | "the cat is on the mat" becomes "cat mat" |
| `remove_function_words` | Drops articles and auxiliaries | "the cat is running" becomes "cat running" |
| `pos_keep_only` | Keeps only content words | "I need to read the book quickly" becomes "need read book" |
| `lemmatize` | Reduces words to root forms | "running studies" becomes "run study" |
| `shorten_synonyms` | Substitutes shorter synonyms | "automobile" becomes "car" |
| `preserve_named_entities` | Protects names from pruning | "New York" stays intact (modifier flag) |
| `normalize_whitespace_punct` | Cleans up spacing | "hello   world!!!" becomes "hello world!" (always on) |

### What never gets removed

Two categories of words are hard-coded as protected, even at the most aggressive setting.

First, negations. Words like `not`, `no`, `never`, `nothing`, `nor`, `nobody`, and `cannot`. Dropping these flips the meaning of a sentence, which would be catastrophic. "Do not run this code" becoming "Do run this code" is not a tradeoff anyone wants.

Second, question words. `What`, `why`, `how`, `when`, `where`, `which`. These carry the intent of a query.

Also, if your original prompt ended with a question mark, the compressed version will too. We re-assert question form at the end of the pipeline so it isn't lost during pruning.

### Four presets you can copy

You don't have to figure out which flags to combine. Here are four named recipes for different aggression levels:

```python
# SAFE: barely shrinks anything, near-perfect quality preservation.
# Useful when you can't afford any quality risk.
compress(prompt,
         remove_filler_phrases=1,
         apply_contractions=1,
         remove_filler_words=1)
# about 2% reduction, 0.96 BERTScore

# BALANCED: the production default. Roughly 30% reduction with minimal
# quality loss. Start here.
compress(prompt,
         remove_filler_phrases=1,
         apply_abbreviations=1,
         apply_contractions=1,
         remove_filler_words=1,
         remove_stopwords=1)
# about 30% reduction, 0.91 BERTScore

# AGGRESSIVE: pure POS-based pruning. Slightly more reduction than balanced
# at very similar quality. Great for high-volume systems.
compress(prompt,
         pos_keep_only=1,
         preserve_named_entities=1)
# about 35% reduction, 0.91 BERTScore

# MAXIMUM: everything on. About 40% reduction at the cost of some output
# quality. Use when the cost savings really matter.
compress(prompt,
         remove_filler_phrases=1, apply_abbreviations=1, apply_contractions=1,
         remove_filler_words=1, remove_stopwords=1, remove_function_words=1,
         pos_keep_only=1, lemmatize=1, shorten_synonyms=1, preserve_named_entities=1)
# about 40% reduction, 0.88 BERTScore
```

## compress_structured: protect the parts that matter

Real prompts are rarely just instructions. They often carry parts that must survive exactly, like a JSON output schema, or rules that would break if a single word were dropped. Compressing those parts the same way you compress the instruction body is dangerous.

`compress_structured()` solves this by letting you assign a compression *level* to each part of the prompt:

| Level | What happens | Use it for |
|-------|--------------|------------|
| `free` | Full compression using your chosen flags | The instruction body |
| `careful` | Only safe, meaning-preserving techniques (no stopword removal, no pruning, no synonyms) | Rules and constraints |
| `protected` | Returned byte-for-byte, untouched | JSON schemas, output formats, examples |

### The easy way: name your sections

The most common case is an instruction, some rules, and an output format. Just pass them as named arguments. The compression flags you pass apply only to the instruction.

```python
from less_tokens import compress_structured

prompt = compress_structured(
    instruction="I was wondering if you could analyse this customer review and tell me how the person is feeling about the product.",
    rules="Do not include any personal opinions. Never guess if you are unsure.",
    output_format='{"sentiment": "positive|negative|neutral", "confidence": 0.0-1.0}',
    remove_stopwords=1,
    remove_filler_phrases=1,
)

print(prompt)
```

Output:

```
analyse customer review tell person feeling product.

don't include any personal opinions. Never guess if you're unsure.

Output format:
{"sentiment": "positive|negative|neutral", "confidence": 0.0-1.0}
```

Look at what happened to each part:

- The **instruction** got compressed hard. "I was wondering if you could" is gone, stopwords are gone.
- The **rules** were compressed gently. "Do not" became "don't" and "you are" became "you're", but the critical words "not" and "Never" survived intact. The meaning is identical.
- The **output format** is byte-for-byte unchanged. Your JSON schema is safe.

### The flexible way: explicit zones

If you need full control over ordering, or you want to mix levels in a custom way, pass an explicit list of zones. Each zone is a dict with `text` and `level`, or a simple `(text, level)` tuple.

```python
from less_tokens import compress_structured

prompt = compress_structured(zones=[
    {"text": "I was wondering if you could summarize the following article.", "level": "free"},
    {"text": "Do not exceed 100 words. Never add facts not in the source.",   "level": "careful"},
    {"text": '{"summary": "...", "word_count": N}',                           "level": "protected"},
])
```

### Why "careful" mode exists

This is the most important design decision in the library. Rules carry meaning in their small words. If you ran full stopword removal on "Do not exceed 100 words" you might get "exceed 100 words", which is the exact opposite instruction. So careful mode disables every technique that could flip or blur meaning:

| Technique | free | careful | Why careful skips it |
|-----------|:----:|:-------:|----------------------|
| Filler phrase removal | yes | yes | Safe, only removes hedging |
| Contractions | yes | yes | Safe, "do not" to "don't" keeps meaning |
| Filler word removal | yes | yes | Safe, "basically" carries no logic |
| Stopword removal | yes | no | Can drop words that matter in a constraint |
| Function word pruning | yes | no | Can drop "not", "all", "only" type logic |
| POS-keep | yes | no | Too aggressive for precise rules |
| Lemmatize | yes | no | Can blur tense or number that matters |
| Synonym shortening | yes | no | Can pick a narrower or wrong synonym |

If even careful mode feels too risky for a specific rule, mark it `protected` and it won't be touched at all.

### Seeing what changed

Pass `return_detail=True` to get a breakdown of every zone, useful for debugging:

```python
result = compress_structured(
    instruction="Please analyse this in detail.",
    output_format='{"x": 1}',
    remove_stopwords=1,
    return_detail=True,
)

print(result["compressed"])     # the assembled prompt
for zone in result["zones"]:
    print(zone["level"], zone["original_len"], "->", zone["compressed_len"])
```

## compare: measure the quality tradeoff

Compression is only useful if the LLM still produces the same answer. `compare()` quantifies that across six different similarity metrics so you can see exactly what compressing cost you.

You make the LLM calls yourself, with whichever provider you like. `compare()` only looks at the four strings: original prompt, compressed prompt, output from original, output from compressed.

```python
from less_tokens import compress, compare
from openai import OpenAI

client = OpenAI()

def call_llm(prompt: str) -> str:
    r = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": prompt}],
        temperature=0,
    )
    return r.choices[0].message.content

original   = "I was wondering if you could explain how to brew good coffee at home."
compressed = compress(original, remove_filler_phrases=1, remove_stopwords=1)

out_original   = call_llm(original)
out_compressed = call_llm(compressed)

metrics = compare(original, compressed, out_original, out_compressed)
```

### What you get back

```python
{
    "compression": {
        "original_tokens":     18,
        "compressed_tokens":   8,
        "token_reduction_pct": 55.56,    # you saved 55% of your tokens
        "original_chars":      72,
        "compressed_chars":    32,
        "char_reduction_pct":  55.56,
    },
    "prompt_similarity": {
        "cosine": 0.842,                 # the two prompts mean roughly the same thing
    },
    "output_similarity": {               # six metrics on the LLM outputs
        "cosine":      0.917,
        "bleu":        0.412,
        "rouge1_f":    0.673,
        "rouge2_f":    0.418,
        "rougeL_f":    0.601,
        "bertscore_p": 0.923,
        "bertscore_r": 0.918,
        "bertscore_f": 0.920,
    },
}
```

### What each of the six metrics actually means

All six measure the same thing from different angles: how similar is the LLM's response to the compressed prompt, compared to its response to the original. Each one captures a different notion of "similar".

**1. `cosine`. Semantic similarity. Range 0.0 to 1.0.**

The plain-English question it answers: *do the two outputs mean the same thing?*

It works by embedding both outputs with SentenceBERT (MiniLM-L6-v2) and taking the cosine of the angle between them. This is the most forgiving metric in the set because it handles paraphrasing well.

Interpretation:
- 0.95 or above: essentially identical meaning
- 0.85 to 0.95: same meaning, different wording
- 0.70 to 0.85: related but starting to drift
- below 0.70: the meanings have meaningfully diverged

**2. `bleu`. Word-sequence overlap. Range 0.0 to 1.0.**

The plain-English question: *do the two outputs use the same exact words in the same order?*

BLEU-4 with smoothing, originally invented for machine translation (Papineni et al., 2002). This is very strict. It penalises rewording, even when the meaning is preserved perfectly.

Interpretation:
- 0.50 or above: near-identical phrasing
- 0.20 to 0.50: similar content but reworded
- below 0.20: very different word choices (which doesn't mean the answer is wrong, just that the LLM phrased it differently)

Don't panic if BLEU is low. That's expected when an LLM rephrases the same answer using different words.

**3. `rouge1_f`. Single-word overlap. Range 0.0 to 1.0.**

The plain-English question: *do the two outputs use the same words, regardless of order?*

ROUGE-1 F1 (Lin, 2004). Measures unigram overlap. Less strict than BLEU because word order doesn't matter.

Interpretation:
- 0.70 or above: strong vocabulary overlap
- 0.40 to 0.70: moderate overlap
- below 0.40: mostly different vocabulary

**4. `rouge2_f`. Two-word phrase overlap. Range 0.0 to 1.0.**

The plain-English question: *do the two outputs share the same two-word phrases?*

ROUGE-2 F1. Same idea as ROUGE-1 but measures bigrams (consecutive word pairs). Stricter than ROUGE-1 because the words have to appear in the same order locally.

Interpretation:
- 0.40 or above: strong phrasal similarity
- 0.15 to 0.40: some shared phrases
- below 0.15: mostly different phrasing

**5. `rougeL_f`. Longest matching subsequence. Range 0.0 to 1.0.**

The plain-English question: *what's the longest stretch of words that appear in both outputs in the same order?*

ROUGE-L F1. Measures the longest common subsequence: words that appear in both outputs in the same order, but allowing other words between them. Captures structural similarity better than BLEU does.

Interpretation:
- 0.60 or above: strong structural alignment
- 0.30 to 0.60: some shared structure
- below 0.30: mostly independent structure

**6. `bertscore_f`. Contextual semantic similarity. Range 0.0 to 1.0.**

The plain-English question: *do the two outputs convey the same ideas, accounting for context?*

BERTScore F1 (Zhang et al., 2020). Computes per-token cosine similarity in a BERT embedding space, matching each token in one output to its most similar token in the other. This is the headline quality metric and correlates better with human judgment than any of the metrics above.

Interpretation:
- 0.95 or above: essentially equivalent outputs
- 0.90 to 0.95: very close, with some phrasing differences
- 0.85 to 0.90: similar core content but noticeable rewording
- below 0.85: meaningful divergence

BERTScore also gives you `bertscore_p` for precision and `bertscore_r` for recall. F1 is the harmonic mean of both, and is the one you should focus on.

### Which metric should you care about?

It depends what you're trying to measure:

| Use case | Look at this | Threshold to aim for |
|----------|--------------|----------------------|
| General quality check | `bertscore_f` | 0.90 or higher |
| You need exact specific words in the output | `bleu` | 0.40 or higher |
| You need the same vocabulary, word order flexible | `rouge1_f` | 0.60 or higher |
| Cheap sanity check without downloading BERT model | `cosine` | 0.85 or higher |

If you don't want the 1 GB BERTScore model downloaded, skip it:

```python
metrics = compare(original, compressed, out_original, out_compressed,
                  bertscore=False)
```

You still get the other five metrics, which together are very informative.

## Async support

Compression is CPU-bound and pure Python, so the async functions run it in a thread executor and never block your event loop. They take exactly the same arguments as their synchronous counterparts.

```python
import asyncio
from less_tokens import acompress, acompress_structured

async def main():
    # Async version of compress()
    short = await acompress(
        "I was wondering if you could help me with this",
        remove_filler_phrases=1, remove_stopwords=1,
    )

    # Async version of compress_structured()
    prompt = await acompress_structured(
        instruction="Please analyse this text in detail.",
        output_format='{"result": "..."}',
        remove_stopwords=1,
    )

    # Compress many prompts at once
    results = await asyncio.gather(
        acompress(p1, remove_stopwords=1),
        acompress(p2, remove_stopwords=1),
        acompress(p3, remove_stopwords=1),
    )

asyncio.run(main())
```

This is handy when you're compressing inside an async web server (FastAPI, aiohttp) or processing a large batch of prompts concurrently.

## A complete example

Here's the whole flow end to end, using structured compression to protect an output format:

```python
from less_tokens import compress_structured, compare
from openai import OpenAI

client = OpenAI()

def ask_gpt(prompt: str) -> str:
    r = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": prompt}],
        temperature=0,
    )
    return r.choices[0].message.content

# Build the original prompt the long way
original = (
    "I was wondering if you could please analyse the following customer "
    "review and tell me the overall sentiment.\n\n"
    "Do not include any personal opinions. Never guess if you are unsure.\n\n"
    'Output format:\n{"sentiment": "positive|negative|neutral", "confidence": 0.0-1.0}'
)

# Compress it, protecting the rules and output format
compressed = compress_structured(
    instruction="I was wondering if you could please analyse the following customer review and tell me the overall sentiment.",
    rules="Do not include any personal opinions. Never guess if you are unsure.",
    output_format='{"sentiment": "positive|negative|neutral", "confidence": 0.0-1.0}',
    remove_filler_phrases=1,
    remove_stopwords=1,
)

print(f"Original   ({len(original)} chars)")
print(f"Compressed ({len(compressed)} chars)")
print()

out_original   = ask_gpt(original)
out_compressed = ask_gpt(compressed)

metrics = compare(original, compressed, out_original, out_compressed)

print(f"Token reduction: {metrics['compression']['token_reduction_pct']}%")
print(f"BERTScore F1:    {metrics['output_similarity']['bertscore_f']}")
```

You shrink the wordy instruction, keep the rules safe, keep the JSON schema exact, and confirm with `compare()` that the model still returns the same structured answer.

## Under the hood

`less-tokens` is built on classical lexical NLP. These are the same techniques used in information retrieval and pre-neural NLP pipelines, just packaged together with sensible defaults and safety guarantees:

- **NLTK** (Loper and Bird, 2002) handles tokenisation, POS tagging, and named entity recognition
- **WordNet** (Miller, 1995) provides the synonym graph
- **tiktoken** counts tokens the same way GPT models do
- **sentence-transformers** computes cosine similarity
- **bert_score** computes BERTScore F1
- **rouge_score** computes ROUGE-1, ROUGE-2, and ROUGE-L
- **NLTK's BLEU** with method-1 smoothing

Every technique is a pure function. Same input plus same flags always produces the same output, byte for byte. Compression itself runs in well under 100 ms on a single CPU core.

## Limitations

A few honest caveats so you know what you're getting.

English only. NLTK stopwords and WordNet are English-language. Multilingual support is open work.

Best on short and medium prompts. Roughly 60 to 2000 characters. Very long retrieval-augmented contexts aren't the target use case. For those, look at learned compressors like [LLMLingua](https://github.com/microsoft/LLMLingua).

The `shorten_synonyms` flag is the riskiest. WordNet sometimes picks topically narrower terms. Don't enable it without testing on your own data first.

Quality is task-dependent. Open-ended Q&A and creative writing tolerate compression well. Commonsense reasoning (HellaSwag-style multiple choice) degrades faster.

`compare()` measures similarity, not correctness. If your original prompt produces a bad LLM output, a similar compressed output is still bad. Make sure your prompts work first, then compress.

## Contributing

Issues and pull requests are very welcome at [github.com/shaminchokshi/less-tokens](https://github.com/shaminchokshi/less-tokens).

To run the test suite locally:

```bash
git clone https://github.com/shaminchokshi/less-tokens.git
cd less-tokens
pip install -e ".[dev]"
pytest tests/ -v
```

## License

MIT. See [LICENSE](https://github.com/shaminchokshi/less-tokens/blob/main/LICENSE).

## Citations

If you're using `less-tokens` in research, the underlying techniques come from these foundational papers:

- **NLTK**: Loper and Bird (2002). *NLTK: The Natural Language Toolkit.* ACL Workshop.
- **WordNet**: Miller (1995). *WordNet: A Lexical Database for English.* CACM 38(11).
- **BERTScore**: Zhang et al. (2020). *BERTScore: Evaluating Text Generation with BERT.* ICLR.
- **BLEU**: Papineni et al. (2002). *BLEU: a Method for Automatic Evaluation of Machine Translation.* ACL.
- **ROUGE**: Lin (2004). *ROUGE: A Package for Automatic Evaluation of Summaries.* ACL Workshop.
- **Sentence-BERT**: Reimers and Gurevych (2019). *Sentence-BERT.* EMNLP.

Related work on prompt compression you might want to compare against:

- **LLMLingua**: Jiang et al. (2023). EMNLP. Learned token pruning with an auxiliary LM, up to 20x compression.
- **Selective Context**: Li et al. (2023). EMNLP. Self-information-based pruning.
