less-tokens 0.5.0

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 document-to-markdown reduction for PDFs and Word files, zone-aware compression that protects JSON schemas and output formats, async support, and a built-in 6-metric quality evaluator.

less-tokens

Shrink your LLM prompts by 30 to 40 percent without changing what the model says back.

less-tokens is a small Python library for developers who are paying for tokens and want to stop paying for the ones that don't earn their place. It compresses prompts before you send them to an LLM, stripping out filler words, redundant phrases, and grammatical scaffolding 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 — same input, same flags, same output, every time. That matters when you're putting something in a production pipeline.

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
• Install
• The functions at a glance
• compress: shrink a prompt
• reduce_document: turn a file into clean markdown
• compress_structured: protect the parts that matter
• smart_compress: compress a conversation message
• compare: measure the quality tradeoff
• Async support
• A complete example
• Under the hood
• Limitations

Why this exists

If you're calling OpenAI, Anthropic, or any other LLM API at meaningful volume, every token is a line item on your bill. And the prompts your code sends carry a lot of fat that the model quietly ignores:

• "I was wondering if you could..." is hedging that adds nothing
• "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

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

There's a second source of waste that bites you the moment your use case involves files. When your pipeline hands an LLM a raw PDF or Word document, you're shipping embedded fonts, positioning data, and office XML on top of the words you actually care about. If your use case only needs the content of the file, converting it to Markdown first cuts the token count enormously — that's what reduce_document() is for.

Install

pip install less-tokens

That single command pulls in everything: the compressor, the compare() metrics stack, and the PDF/Word parsers used by reduce_document(). There are no optional extras to remember and nothing else to wire up.

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:

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 five functions. Pick the one that matches what your use case actually needs:

Function	Reach for it when…
compress()	You have a prompt string and want it shorter
compress_structured()	Your prompt mixes free instructions with parts you can't touch, like a JSON output schema or strict rules
reduce_document()	Your input is a PDF or Word file and you only need its content as text, not a full file upload
smart_compress()	You have a single conversation message (user or LLM) that mixes prose and code/tables/URLs and you want only the prose compressed
compare()	You want to prove the compression didn't change the model's answer
acompress() / acompress_structured() / areduce_document() / asmart_compress()	You're doing any of the above inside an async event loop

The mental model: if you start with a string, use compress() (or compress_structured() if parts are sacred). If you start with a file, run reduce_document() first to get text, then optionally compress() that. If you're compressing a conversation history, use smart_compress() on each message. When you want to know what it cost you in quality, run compare().

compress: shrink a prompt

This is the workhorse. 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 opt in to exactly the behavior your use case can tolerate.

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, because dropping them would silently corrupt the instruction your code is sending.

First, negations. Words like not, no, never, nothing, nor, nobody, and cannot. Dropping these flips the meaning of a sentence, which would be catastrophic in a production prompt. "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 mapped to how much risk your use case can absorb:

# SAFE: barely shrinks anything, near-perfect quality preservation.
# Use it when you can't afford any quality risk at all.
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 where cost dominates.
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 savings genuinely outweigh the quality hit.
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

reduce_document: turn a file into clean markdown

If your AI use case only requires the content of a PDF or a Word file — and not an entire multimodal text-plus-file upload — don't hand the raw file to the model. A raw .pdf or .docx is mostly not content: it's embedded fonts, per-glyph positioning, style definitions, office XML, page geometry. The model doesn't need any of that, but every byte of it costs you tokens.

Scrape the content instead. reduce_document() strips all that unnecessary info — the layout details, the metadata, the fonts and spacing — and keeps only the parts that actually carry meaning: titles, headings, bullet and numbered lists, tables. The result is far fewer tokens, and what you get back is clean Markdown the model reads happily.

And guess what: if your use case permits you to go even leaner, you can run that Markdown straight through compress() and shrink it again. File → clean Markdown → compressed text, each step cheaper than the last.

Basic usage

from less_tokens import reduce_document

markdown = reduce_document("quarterly_report.pdf")
print(markdown)

# Quarterly Report

## Summary

Revenue grew 18% quarter over quarter, driven mainly by the new
enterprise tier.

## Key figures

| Metric   | Q2    | Q3    |
| ---      | ---   | ---   |
| Revenue  | 4.1M  | 4.8M  |
| Churn    | 2.3%  | 1.9%  |

## Next steps

- Expand the sales team
- Launch in two new regions

Drop that Markdown straight into a prompt, store it, or compress it further. It's just text now.

Parameters

Parameter	What it does
path	Path to the document. PDF, Word, or any plain-text format.
file_type	Force a parser regardless of extension, e.g. "pdf" or ".docx". Handy when your pipeline receives files with missing or wrong extensions.
include_tables	True by default. Converts tables to Markdown tables. Set False to skip table detection entirely.

What it keeps and what it drops

Kept (the content your model needs)	Dropped (the overhead you were paying for)
Titles and headings (as #, ##, ...)	Margins, indentation, page size
Paragraph text	Fonts, font sizes, colors
Bullet and numbered lists	Line and paragraph spacing
Tables (as Markdown tables)	Absolute positioning, page geometry
Bold and italic emphasis	Headers, footers, page numbers
Reading order	Office XML and style definitions

Supported file types

Type	Extensions
PDF	.pdf
Word	.docx, .docm
Plain text / Markdown	.txt, .md, .rst, ...

All of these work out of the box with a plain pip install less-tokens — the PDF and Word parsers ship as part of the package.

Pairing it with compress

This is the two-step move that gets your file-based use case to the smallest possible footprint: first strip the file down to its content, then compress that content lexically.

from less_tokens import reduce_document, compress

# Step 1: file -> clean markdown (drops layout + metadata)
content = reduce_document("contract.docx")

# Step 2: markdown -> compressed text (drops filler + stopwords)
lean = compress(content,
                remove_filler_phrases=1,
                remove_stopwords=1,
                apply_contractions=1)

# `lean` is now a tiny fraction of the original file's token count.

One caution worth building into your code: if the document has tables you need intact, aggressive compress() flags (stopword removal, POS-keep) will chew up the cell text and pipe structure. Either keep reduce_document(..., include_tables=False) if you don't need them, or protect the table with compress_structured() (next section).

compress_structured: protect the parts that matter

The real prompts your application builds are rarely just instructions. They carry parts that must survive exactly — a JSON output schema, an example the model copies, or rules that break if a single word is dropped. Compressing those parts the same way you compress the instruction body will quietly corrupt your output contract.

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 in real code 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.

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, so your parser downstream won't break.

The flexible way: explicit zones

When 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.

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, and the one that keeps it safe to drop into production. 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 when you're debugging why an output contract broke:

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"])

smart_compress: compress a conversation message

When you are working with a multi-turn conversation history — a list of user inputs and LLM responses — you cannot run compress() directly on each message. LLM responses routinely mix natural language prose with elements that must never be touched: fenced code blocks, inline code, Markdown tables, URLs, math expressions, and HTML. Compressing those would corrupt the code or break the output contract.

smart_compress() solves this. It parses each message, automatically detects every protected zone, and compresses only the natural language prose around them. Apply it to every message in your history:

from less_tokens import smart_compress

compressed_history = [
    smart_compress(msg, remove_filler_phrases=1, remove_stopwords=1)
    for msg in conversation
]

What is protected and what is compressed

Protected (returned verbatim)	Compressed (natural language only)
Fenced code blocks (```)	Paragraph prose
Indented code blocks	Heading text (the # prefix is kept)
Inline code (`backticks`)	List-item text (the - / 1. marker is kept)
Markdown tables
Bare URLs and Markdown links
Math blocks ($$...$$) and inline math ($...$)
HTML tags
JSON / array blocks

Debugging with return_segments

Pass return_segments=True to see exactly what was protected and what was compressed:

result = smart_compress(msg, remove_filler_phrases=1, return_segments=True)
print(result["compressed"])
for seg in result["segments"]:
    print(seg["kind"], "->", seg["original"][:40])

compare: measure the quality tradeoff

Compression is only worth shipping if the LLM still produces the answer your use case depends on. compare() quantifies that across six different similarity metrics, so you can decide based on numbers instead of vibes.

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

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

{
    "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", and which one you care about depends on what your use case promises its users.

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 on what your use case is actually promising:

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 your environment can't afford the 1 GB BERTScore model download, skip it:

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

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

Async support

If your use case runs inside an async web server or processes prompts in large concurrent batches, the async functions run the (CPU-bound, pure-Python) work in a thread executor so they never block your event loop. They take exactly the same arguments as their synchronous counterparts.

Sync	Async
compress()	acompress()
compress_structured()	acompress_structured()
reduce_document()	areduce_document()
smart_compress()	asmart_compress()

import asyncio
from less_tokens import acompress, acompress_structured, areduce_document, asmart_compress

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,
    )

    # Async version of reduce_document()
    content = await areduce_document("report.pdf")

    # Async version of smart_compress() — compress a full conversation history concurrently
    compressed_history = await asyncio.gather(
        *[asmart_compress(msg, remove_filler_phrases=1, remove_stopwords=1)
          for msg in conversation]
    )

    # Or reduce a batch of uploaded files at once
    docs = await asyncio.gather(
        areduce_document("a.pdf"),
        areduce_document("b.docx"),
        areduce_document("c.txt"),
    )

asyncio.run(main())

This is what you want when you're compressing inside FastAPI or aiohttp, or reducing a batch of user-uploaded files concurrently.

A complete example

Here's the whole flow end to end for a file-based use case: a user uploads a review as a PDF, you pull out just the content, compress the wordy instruction, protect the output schema, and verify with compare() that the model still returns the same structured answer your code depends on.

from less_tokens import reduce_document, 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

# Step 0: a user uploaded a review as a PDF. Pull out just the content.
review = reduce_document("customer_review.pdf")

# Build the original prompt the long way
original = (
    "I was wondering if you could please analyse the following customer "
    f"review and tell me the overall sentiment.\n\n{review}\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(zones=[
    ("I was wondering if you could please analyse the following customer "
     f"review and tell me the overall sentiment.\n\n{review}", "free"),
    ("Do not include any personal opinions. Never guess if you are unsure.", "careful"),
    ('{"sentiment": "positive|negative|neutral", "confidence": 0.0-1.0}', "protected"),
],
    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 pulled the content out of the file, shrank the wordy instruction, kept the rules safe, kept the JSON schema exact, and confirmed with compare() that the model still returns the same structured answer. That's the full library working together on one realistic use case.

Under the hood

less-tokens is built on classical lexical NLP — the same techniques used in information retrieval and pre-neural NLP pipelines, packaged together with sensible defaults and safety guarantees so you can drop them into real code:

• 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
• PyMuPDF gives us the raw text spans (with font size and bold/italic flags) and table regions of a PDF; reduce_document() reconstructs the Markdown from those primitives itself — headings from relative font size, emphasis from span flags, lists from leading glyphs, and reading order from on-page position
• python-docx reads Word documents in true reading order, which reduce_document() maps to Markdown headings, lists, and tables

Every compression technique is a pure function. Same input plus same flags always produces the same output, byte for byte — which is exactly what you want when the thing sits in a deterministic pipeline. Compression itself runs in well under 100 ms on a single CPU core, and document reduction is deterministic too: the same file always produces the same Markdown.

Limitations

A few honest caveats so you know whether this fits your use case before you build on it.

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.

The shorten_synonyms flag is the riskiest. WordNet sometimes picks topically narrower terms. Don't enable it in production 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.

reduce_document() reads text, not pixels. Scanned PDFs or image-only documents have no extractable text layer, so they come back empty — run OCR first if that's your input. It also doesn't handle the old binary .doc format (convert to .docx first), and complex multi-column or heavily nested table layouts may not map cleanly onto Markdown.

Contributing

Issues and pull requests are very welcome at github.com/shaminchokshi/less-tokens.

To run the test suite locally:

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

License

MIT. See 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.