Skip to content

h2cker/vecr

Repository files navigation

vecr-compress

English | 中文

Auditable, deterministic context compression for LLMs. Structured data — order IDs, URLs, dates, citations, code — survives compression by an explicit regex whitelist you can inspect, extend, and audit. Every pin and every drop is logged: you get a retained_matches list and a dropped_segments list on every call.

PyPI version Python versions License Tests Downloads

Why this exists

A 2026 Factory.ai production study found that "artifact tracking" (IDs, file paths, error codes) is the worst-compressed category across every compressor tested — scoring just 2.19–2.45 out of 5.0, worse even than OpenAI's native compaction (3.43/5.0). No shipped library offers a deterministic retention primitive: they all rely on LLM judgment or learned scoring that can silently drop a customer ID, a transaction amount, or a compliance citation. vecr-compress solves exactly that gap. It does not claim the highest compression ratio — that is Compresr's lane. It offers an auditable, extensible whitelist-based compressor you can reason about end-to-end.

30-second example

from vecr_compress import compress

messages = [
    {"role": "system", "content": "You are a refund analyst."},
    {"role": "user", "content":
        "Hi there! Thanks so much for reaching out today. "
        "Hope you are having a wonderful morning. "
        "We really appreciate you writing in. "
        "Happy to take a look at this for you. "
        "Totally understand how important this is. "
        "Sure thing, let me pull up the record. "
        "Absolutely, this is our top priority. "
        "The refund request references order ORD-99172 placed on 2026-03-15. "
        "The customer email is buyer@example.com. "
        "The total charge was $1,499.00 on card ending 4242. "
        "Thanks again for your patience and have a great day!"},
    {"role": "user", "content": "What is the order ID and refund amount?"},
]

result = compress(messages, budget_tokens=80, protect_tail=1)

for m in result.messages:
    print(m["role"], "->", m["content"])

print(f"\n{result.original_tokens} -> {result.compressed_tokens} tokens "
      f"({result.ratio:.2%}); pinned {len(result.retained_matches)} facts, "
      f"dropped {len(result.dropped_segments)} segments")

Every structured fact in the input — ORD-99172, 2026-03-15, buyer@example.com, $1,499.00 — survives, because each is pinned by the retention whitelist before the knapsack budget packing runs. Filler sentences like "We really appreciate you writing in" and "Sure thing, let me pull up the record" are dropped. On this fixture: 131 → 78 tokens (~60%), 3 facts pinned, 6 filler segments dropped.

Why protect_tail=1? By default the last two messages are pinned untouched (so the user turn with the actual question is always preserved). Here the penultimate turn is the body we want to compress, so we drop the tail pin to 1. See the API docs for protect_tail / protect_system semantics.

The retention contract

vecr-compress ships 13 built-in rules. Any segment containing a match is pinned — kept regardless of token budget. If total pinned content exceeds the budget, the compressor returns all pinned segments and logs a warning rather than silently dropping facts.

Pattern Example match Why it matters
uuid 3f6e4b1a-23cd-4e5f-9012-abcdef012345 Trace IDs, session IDs, correlation keys
date 2026-03-15, 2026-03-15T09:30:00 Deadlines, timestamps, audit trails
code-id ORD-99172, INV_2024_A, CUST#42 Order, invoice, customer identifiers
email buyer@example.com Contact records, PII audit
url https://api.example.com/v2/orders Endpoints, evidence links, sources
path /var/log/app/error.log, C:/data/report.csv File references, error locations
code-span `raise ValueError(msg)` Inline code in prose
fn-call process_refund(order_id, amount), obj.method(a, b) (code-like identifiers only) Function references in code review
citation [12], [Smith 2023] Academic and legal citations
json-kv "status": "pending_review" Structured payload fields
hash 9f3ab2c4 (8+ hex chars, 2+ digits) Git SHAs, content digests
number $1,499.00, 12.4%, v3.2.1 Amounts, rates, version strings
integer 9172, 99172, 2026 (4+ digits) IDs, reference numbers, years

Extend the contract with your own rules:

import re
from vecr_compress import compress, RetentionRule, DEFAULT_RULES

custom_rules = DEFAULT_RULES.with_extra([
    RetentionRule(name="invoice", pattern=re.compile(r"INV-\d{6}")),
])
result = compress(messages, budget_tokens=2000, retention_rules=custom_rules)

Details on testing and extending rules: see RETENTION.md.

Benchmark (reproducible)

Needle-in-haystack survival: 11 needles × 3 positions × 6 ratios × 3 configs = 594 trials (bench/needle.py).

Structured needles (7) — baseline vs. L2 retention

ratio baseline + L2 retention
1.00 100% 100%
0.50 100% 100%
0.30 100% 100%
0.15 100% 100%
0.08 100% 100%
0.04 100% 100%

The baseline heuristic scorer keeps all structured tokens in this synthetic fixture. L2 turns that measurement into a deterministic contract — the same 100% holds across any workload, scorer, or distribution, not just this fixture. If ORD-\d+ appears in the input, it will appear in the output.

Stealth needles (4, plain prose) — where the tradeoff shows

ratio baseline + L2 retention
1.00 100% 100%
0.50 100% 100%
0.30 83% 83%
0.15 75% 67%
0.08 75% 0%
0.04 75% 0%

L2's cost: must-keep structured content pins the budget, leaving little room for plain-prose stealth needles at aggressive ratios (target 0.15 → actual 0.16 because the whitelist overrides the budget). On natural-language QA (HotpotQA probe, N=100) a blended question-aware scorer adds +9.9pp supporting-fact survival at ratio 0.5 over L2 alone — opt in with compress(..., use_question_relevance=True) (v0.1.3+). Off by default so the deterministic contract stays loud; worth turning on when your context is long prose and you have a real question. See docs/BENCHMARK.md for details.

Note: actual compression ratio may exceed the target when must-keep content is large — this is intentional and honest behaviour, not a bug.

To reproduce:

pip install -e .
python -m bench.needle

Install

pip install vecr-compress                  # core only (requires tiktoken)
pip install vecr-compress[langchain]       # + LangChain adapter
pip install vecr-compress[llamaindex]      # + LlamaIndex adapter

Requires Python 3.10+.

LangChain / LlamaIndex

Framework adapters are opt-in via extras ([langchain], [llamaindex]). Core compression has no framework dependency.

LangChain — compress a chat history before passing it to any chat model:

from langchain.messages import HumanMessage, SystemMessage
from vecr_compress.adapters.langchain import VecrContextCompressor

compressor = VecrContextCompressor(budget_tokens=2000)
compressed = compressor.compress_messages([
    SystemMessage(content="You are a helpful assistant."),
    HumanMessage(content="Long conversation history..."),
    HumanMessage(content="The actual question."),
])

# NL-QA workload with a real user question? Opt in to question-aware blending:
#   VecrContextCompressor(budget_tokens=2000, use_question_relevance=True)

LlamaIndex — postprocess retrieved nodes before synthesis:

from llama_index.core.schema import NodeWithScore, TextNode
from vecr_compress.adapters.llamaindex import VecrNodePostprocessor

processor = VecrNodePostprocessor(budget_tokens=1500)
kept = processor.postprocess_nodes(nodes, query_str="the user's question")

How it works (30-second tour)

Two layers applied in order:

  1. Retention whitelist — segments matching any built-in rule are pinned and bypass the budget knapsack entirely.
  2. Heuristic packing — remaining segments are scored by entropy and structural signal (digits, braces, capitalization); filler lines like Hi!, Thanks!, As an AI… score 0.0 and are dropped before any budget math; the rest are packed greedily into the token budget.

Question-aware blending is opt-in via compress(..., use_question_relevance=True) (v0.1.3+); it layers a Jaccard overlap score (0.4 weight) on top of the heuristic (0.6 weight). Advanced callers can also pass a custom ScorerFnblended_score, heuristic_score, and question_relevance are all exported from vecr_compress. See RETENTION.md for details.

vs. alternatives

Approach Open source Retention contract
Compresr (YC W26) LLM summarization, hosted model No None — JSON atomic treatment is planned
LLMLingua-2 Probabilistic token classifier Yes None
LangChain DeepAgents compact Autonomous agent-triggered Yes (LangChain) None
Provider-native compaction (OpenAI/Google) Opaque, single-provider No None
vecr-compress Regex whitelist + heuristic knapsack Yes Deterministic, auditable

Choose Compresr for maximum compression ratio. Choose LLMLingua-2 for pure-Python research. Choose vecr-compress when you want an auditable, extensible whitelist-based compressor you can reason about end-to-end, and you can live with v0.1 limits (Python-only, sentence-level granularity, no streaming).

What this does NOT do

  • No streaming. compress() is synchronous and one-shot.
  • No tool-call rewriting. tool_use / tool_result blocks pass through verbatim — safe, zero gain.
  • Sentence-level granularity only. No token-level pruning or learned rewrites.
  • English-tuned. Stopword list and regex patterns are English-first. Multilingual quality is untested.
  • No embedding scorer. Jaccard overlap is lexical. Semantic relevance scoring lives in the reference gateway.

Contributing / License / Links

Apache 2.0. Contributions welcome via the main repo.

About

Deterministic, auditable LLM context compression — regex whitelist guarantees structured facts (IDs, URLs, dates, code) survive. Two layers: retention + heuristic knapsack.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

No contributors

Languages