Compression

Laghav strips filler words, preambles, duplicate lines, and verbose code comments using 8 specialized rules — reducing token usage by an average of 61% without sacrificing response quality.

The 8 compression rules

Rule name	What it strips	Avg saving
`filler`	Politeness hedges ('Hey I wanted to ask you...', 'Could you possibly...')	10–20%
`preamble`	Verbose intro sentences before the actual question	8–15%
`dedup`	Repeated identical or near-identical sentences	5–30%
`intent`	Redundant explanation of what the user wants	5–12%
`whitespace`	Extra blank lines, trailing spaces, redundant newlines	2–8%
`log_slicer`	INFO/DEBUG log spam — keeps only ERROR, WARN + 2-line context	70–98%
`code_comment`	JSDoc, Python docstrings, inline comments (preserves signatures & logic)	15–40%
`json_slim`	Redundant JSON keys, default null values, empty arrays	20–50%

Aggressiveness control

The max_aggressiveness parameter (0.0–1.0) controls how aggressively rules apply. Higher values save more tokens but may reduce quality score. The quality scorer always checks the result before sending.

aggressiveness.py

# Light — only strip obvious filler and whitespace (preserves more context)
response = client.complete(
    messages=messages,
    model="auto",
    laghav_options={"max_aggressiveness": 0.2}
)
 
# Default — balanced compression
response = client.complete(
    messages=messages,
    model="auto",
    laghav_options={"max_aggressiveness": 0.5}
)
 
# Aggressive — maximum token reduction (good for log analysis, JSON payloads)
response = client.complete(
    messages=messages,
    model="auto",
    laghav_options={"max_aggressiveness": 0.9}
)

Content type hints

Laghav auto-detects content type. You can override with the playground API or via content_type in the playground endpoint:

Content type	Preferred rules	Use case
`auto`	All rules, ranked by signal	Default — Laghav detects type
`code`	code_comment, whitespace, dedup	Source code, function context
`log`	log_slicer, whitespace, dedup	Log files, agent traces
`json`	json_slim, whitespace, dedup	API payloads, config files
`text`	filler, preamble, intent, whitespace	Natural language prompts

Skipping specific rules

skip_rules.py

# Skip intent stripping — useful if intent IS the important part
response = client.complete(
    messages=messages,
    model="auto",
    laghav_options={
        "skip_rules": ["intent", "preamble"]
    }
)

Real example

Before (63 tokens)

Hey I just wanted to ask you if you could possibly help me understand the main causes of the revenue drop that happened last quarter. I think it would be really helpful if you could explain it clearly.

After (24 tokens)

Explain main causes of last quarter's revenue drop.

62% compression94/100 quality

ℹOpen source compression core

The 8 compression rules are open source at github.com/laghav-ai/compress (MIT license). Install standalone with pip install laghav-compress for local compression without a Laghav account.

Supported Models Model Routing