tokenpruner

Slash LLM input tokens by 70–80% without losing meaning.

tokenpruner compresses prompts, code context, and multi-turn conversations before they are sent to Claude, GPT-4, or any LLM — reducing cost and latency with zero external model dependencies.

from tokenpruner import TextPruner, PruningConfig, PruningStrategy

pruner = TextPruner(PruningConfig(strategy=PruningStrategy.COMPOSITE))
result = pruner.prune(my_12000_token_prompt)
print(result)
# PruningResult(original=12400tok, pruned=2700tok, saved=78%)

Why tokenpruner?

Installation

pip install tokenpruner

# Optional: exact token counts via tiktoken
pip install "tokenpruner[tiktoken]"

Quick start

Compress a single prompt

from tokenpruner import TextPruner, PruningConfig, PruningStrategy

# COMPOSITE runs: template_strip → code_minify → dedup → semantic → sliding_window
pruner = TextPruner(PruningConfig(
    strategy=PruningStrategy.COMPOSITE,
    target_ratio=0.25,   # keep 25% of tokens = 75% reduction
))

result = pruner.prune(long_prompt)
print(f"Saved {result.reduction_ratio:.0%} — {result.tokens_saved} tokens")

Prune a full conversation

from tokenpruner import ConversationPruner, Message, PruningConfig

msgs = [Message(role=m["role"], content=m["content"]) for m in conversation]

pruner = ConversationPruner(
    PruningConfig(max_tokens=8000),
    keep_recent_turns=3,  # last 3 exchanges verbatim
)
result = pruner.prune(msgs)
pruned_dicts = [{"role": m.role, "content": m.content} for m in result.pruned_messages]

Drop-in Claude adapter

import anthropic
from tokenpruner import PruningConfig, PruningStrategy
from tokenpruner.adapters.claude import ClaudeAdapter

client = anthropic.Anthropic()
adapter = ClaudeAdapter(client, config=PruningConfig(max_tokens=8000))

response, meta = adapter.messages_create(
    model="claude-opus-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": giant_codebase_dump}],
    system="You are a code reviewer.",
)
print(f"Tokens saved: {meta['messages_reduction']:.0%}")

Drop-in OpenAI adapter

import openai
from tokenpruner import PruningConfig
from tokenpruner.adapters.openai import OpenAIAdapter

client = openai.OpenAI()
adapter = OpenAIAdapter(client, config=PruningConfig(max_tokens=6000))

response, meta = adapter.chat_completions_create(
    model="gpt-4o",
    messages=[{"role": "user", "content": long_context}],
)

Strategies

Advanced features

Smart Cache (LRU + TTL)

from tokenpruner import SmartCache, TextPruner

cache: SmartCache = SmartCache(maxsize=256, ttl=300)
pruner = TextPruner()

@cache.memoize
def cached_prune(text: str):
    return pruner.prune(text)

cached_prune(prompt)   # computes
cached_prune(prompt)   # cache hit — free
print(cache.stats())   # {'hits': 1, 'misses': 1, 'hit_rate': 0.5, ...}

Pipeline with audit log and retry

from tokenpruner import PruningPipeline, PruningConfig, PruningStrategy, TextPruner

make_pruner = lambda s: lambda t: TextPruner(PruningConfig(strategy=s)).prune(t).pruned_text

pipeline = (
    PruningPipeline()
    .add_step("dedup", make_pruner(PruningStrategy.DEDUP))
    .add_step("semantic", make_pruner(PruningStrategy.SEMANTIC))
    .with_retry(n=2, backoff=0.5)
)

result, audit = pipeline.run(long_text)
print(audit)  # per-step token counts, duration, errors

# Async
result, audit = await pipeline.arun(long_text)

Declarative validator

from tokenpruner import PruningValidator

validator = (
    PruningValidator()
    .require_min_length(10)
    .require_max_tokens(100_000)
    .require_no_secrets()
    .add_rule("no_pii", lambda t: "@" not in t, "Email address detected")
)

errors = validator.validate(prompt)   # {} = valid
await validator.avalidate(prompt)     # async version
validator.validate_or_raise(prompt)   # raises ValidationError if invalid

Async batch processing

from tokenpruner import async_batch, batch, PruningConfig

# Sync
results = batch(list_of_1000_prompts, concurrency=16)

# Async
results = await async_batch(list_of_1000_prompts, concurrency=32)

# Per-item errors are isolated — one bad item doesn't abort the batch
for r in results:
    if isinstance(r, Exception):
        print("Item failed:", r)
    else:
        print(f"Saved {r.reduction_ratio:.0%}")

Rate limiter

from tokenpruner import RateLimiter, get_rate_limiter

# Global limiter
limiter = RateLimiter(rate=10, capacity=10)
with limiter:
    result = pruner.prune(text)

async with limiter:
    result = pruner.prune(text)

# Per-key limiting (e.g. per user/API key)
limiter = get_rate_limiter("user-abc", rate=5, capacity=5)
print(limiter.stats())

Streaming for large documents

from tokenpruner import stream_prune, async_stream_prune

# Sync streaming
for chunk_result in stream_prune(huge_document, chunk_size=2000):
    send_to_llm(chunk_result.pruned_text)

# Async streaming (cancellation-safe)
async for chunk_result in async_stream_prune(huge_document, chunk_size=2000):
    await send_to_llm(chunk_result.pruned_text)

Diff engine

from tokenpruner import diff_prune

diff = diff_prune(original_prompt)
print(diff.summary())
# Pruning Summary
#   Original : 12,400 tokens, 312 lines
#   Pruned   : 2,730 tokens, 68 lines
#   Removed  : 9,670 tokens (78.0%)
#   ...

data = diff.to_json()   # machine-readable dict

Circuit breaker

from tokenpruner import CircuitBreaker

cb = CircuitBreaker(failure_threshold=5, reset_timeout=30)

@cb.protect
def call_llm_api(prompt: str) -> str:
    ...  # your API call

# Or inline
result = cb.call(pruner.prune, long_text)

print(cb.stats())
# {'state': 'closed', 'failures': 0, 'total_calls': 42, 'rejected_calls': 0}

Compression techniques

tokenpruner applies these evidence-based techniques:

1. Template stripping — removes You are a helpful assistant and empty XML tags (20–40%)
2. Code minification — strips comments, normalises whitespace (40–60% on code)
3. Jaccard deduplication — removes near-duplicate sentences (30–70%)
4. Heuristic semantic scoring — keeps high-value sentences (keyword density, position, structure)
5. Sliding window — retains a prefix anchor + most-recent suffix
6. Hard truncation — deterministic budget enforcement

Optional: exact token counting

from tokenpruner.utils.tokenizers import count_tokens_exact

# Uses tiktoken if installed, falls back to fast estimate
n = count_tokens_exact("hello world", model="cl100k_base")

License

MIT