langchain-vecr-compress

Drop-in LangChain integration for vecr-compress — the only LLM context compressor with a deterministic retention contract. Before your chat history reaches the model, vecr-compress pins every order ID, URL, date, email, and code reference using an auditable regex whitelist, then packs the remaining budget with the most question-relevant sentences. Structured data never disappears silently; tool calls round-trip intact. This partner package is a thin shim so you can install it with the standard LangChain pattern and get started immediately, with all logic staying in the vecr-compress core.

Install

pip install langchain-vecr-compress

This installs vecr-compress and langchain-core automatically.

30-second example

from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
from langchain_vecr_compress import VecrContextCompressor

compressor = VecrContextCompressor(budget_tokens=120)

history = [
    SystemMessage(content="You are a refund analyst."),
    HumanMessage(content="Hi! I have a question about order ORD-99172."),
    AIMessage(
        content="Let me look that up.",
        tool_calls=[{"id": "c1", "name": "lookup_order", "args": {"order_id": "ORD-99172"}}],
    ),
    HumanMessage(content="The charge was $1,499.00 on 2026-03-15. Please advise."),
    HumanMessage(content="Also, totally just saying hi — hope you're having a great day!"),
]

compressed = compressor.compress_messages(history)

for msg in compressed:
    print(type(msg).__name__, "->", msg.content[:80])

The compressor will drop the filler greeting while preserving ORD-99172, $1,499.00, 2026-03-15, and the AIMessage with its tool_calls list fully intact.

tool_calls round-trip guarantee

AIMessage objects carrying tool_calls are converted to Anthropic-style tool_use content blocks internally. The compressor's skip-mask treats any message containing a tool_use block as must-keep and passes it through verbatim. On the way out, the blocks are converted back to AIMessage(content=..., tool_calls=[...]). This round-trip was audited and tested in vecr-compress v0.1.1.

ai_msg = AIMessage(
    content="searching now",
    tool_calls=[{"id": "t1", "name": "web_search", "args": {"q": "refund policy"}}],
)
[out] = compressor.compress_messages([ai_msg])
assert out.tool_calls[0]["name"] == "web_search"  # always true

Advanced: access compression telemetry

result = compressor.compress_with_report(history)
print(f"{result.original_tokens} -> {result.compressed_tokens} tokens ({result.ratio:.1%})")
print(f"Pinned facts: {len(result.retained_matches)}")
for seg in result.dropped_segments:
    print("dropped:", seg["text"][:60])

Extending the retention contract

import re
from vecr_compress import RetentionRule, DEFAULT_RULES

custom_rules = DEFAULT_RULES.with_extra([
    RetentionRule(name="ticket", pattern=re.compile(r"\bTICKET-\d{4,8}\b")),
])
compressor = VecrContextCompressor(budget_tokens=2000, retention_rules=custom_rules)

Links

• Main repo and full docs: https://github.com/h2cker/vecr
• Retention contract details: RETENTION.md
• Issues: https://github.com/h2cker/vecr/issues

License

Apache 2.0 — see LICENSE.