compresr 2.7.1

Python SDK for Compresr - Intelligent prompt compression service

Compresr Python SDK

Query-aware LLM context compression — reduce LLM API costs by 30-70%.

Install

pip install compresr

Get an API key at compresr.ai → Dashboard → API Keys.

Quick start

from compresr import CompressionClient

client = CompressionClient(api_key="cmp_your_api_key")

result = client.compress(
    context="Long passage to compress...",
    query="What is the main conclusion?",
    target_compression_ratio=0.5,
)

print(f"Original:   {result.data.original_tokens} tokens")
print(f"Compressed: {result.data.compressed_tokens} tokens")
print(f"Saved:      {result.data.tokens_saved} tokens")
print(result.data.compressed_context)

The default model is latte_v1 (query-aware). Pass any other model name your account has access to via compression_model_name="..." — the backend validates.

Batch

Compress up to 100 contexts in one call. Pass a single query (applied to all) or a list of one query per context:

batch = client.compress_batch(
    contexts=["Doc 1...", "Doc 2...", "Doc 3..."],
    queries="What is self-attention?",
    target_compression_ratio=0.5,
)

print(f"Total saved: {batch.data.total_tokens_saved} tokens")

Async + streaming

result = await client.compress_async(context="...", query="...")

for chunk in client.compress_stream(context="...", query="..."):
    print(chunk.content, end="")

LLM-agnostic agent client

One CompressionClient, three provider-shape facades, one engine. Construct the client with llm= and you get an agent surface where every tool output is compressed automatically before the LLM sees it.

import os
from compresr import CompressionClient, WebSearchTool

client = CompressionClient(
    api_key=os.environ["COMPRESR_API_KEY"],
    llm="anthropic",                        # or "openai", "google_genai"
    llm_api_key=os.environ["ANTHROPIC_API_KEY"],
    compression={"target_compression_ratio": 0.5, "min_tokens": 300},
)

Use llm="anthropic:claude-haiku-4-5" if you want a default — but the call-site model= always wins.

Three equivalent surfaces sit on the same client — the model lives at the call site, just like Anthropic's and OpenAI's own SDKs:

# Anthropic shape
client.messages.create(model="claude-haiku-4-5", max_tokens=512,
                       messages=[...], tools=[...])

# OpenAI shape
client.chat.completions.create(model="gpt-4o-mini", messages=[...], tools=[...])

# Native — returns a NormalizedResult
client.run(prompt="...", model="claude-haiku-4-5", tools=[...], max_tokens=512)

Behind all three sits LangChain 1.0's create_agent + CompresrToolMiddleware. Tool outputs above min_tokens flow through client.compress(...) first.

Built-in web search

search = WebSearchTool.tavily(
    api_key=os.environ["TAVILY_API_KEY"],
    max_results=5,
    allowed_domains=["nytimes.com", "reuters.com"],   # optional
)
# Brave: WebSearchTool.brave(api_key=..., max_results=5)

Bring your own tool

Any @tool-decorated function works — its string output is compressed for you:

from langchain_core.tools import tool

@tool
def kb_lookup(topic: str) -> str:
    """Look up the internal policy on the given topic."""
    return INTERNAL_KB.get(topic, "Not found.")

client.messages.create(model="claude-haiku-4-5", max_tokens=256,
                       messages=[{"role": "user", "content": "Refund policy?"}],
                       tools=[kb_lookup])

Switch providers with one line: llm="openai" instead of llm="anthropic" (then pass the model at the call site). Tools and code are unchanged.

Per-call LLM knobs

Pass temperature, top_p, max_tokens, stop_sequences, presence_penalty, frequency_penalty, seed, etc. to any facade — they're forwarded to the underlying chat model via .bind(...) per call, so the cached chat model is never mutated:

client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=512,
    temperature=0.2,
    top_p=0.9,
    messages=[...],
)

Gemini's max_output_tokens is aliased automatically when targeting llm="google_genai:...".

Why not provider-native server search? Anthropic's web_search_20250305, OpenAI's web_search_preview, and Gemini's google_search run server-side and return encrypted/opaque content that Compresr cannot read or compress. Use Tavily or Brave so the result is plaintext we can compress.

Research agent

client.research.run(question) runs a multi-step web-research loop with per-snippet latte_v1 compression on tool results and multi-provider prompt caching. Loop structure adapted from Perplexity search_evals (MIT).

from compresr import CompressionClient

client = CompressionClient(
    api_key="cmp_...",
    llm="anthropic:claude-sonnet-4-6",
    llm_api_key="sk-ant-...",
)

result = client.research.run(
    "What was the latest stable Python version released in 2025?",
    search="tavily",          # "tavily" | "brave" | a LangChain BaseTool
    max_steps=10,
)

print(result.answer)              # parsed Exact Answer field
print(result.explanation)         # parsed Explanation field
print(result.confidence)          # parsed 0-1 confidence
print(result.citations)           # list[Citation(url=...)]
print(result.usage.cache_read_tokens, result.usage.calls)

Single-shot mode: client.research.search(question) runs one search + a forced final answer (equivalent to run(..., max_steps=2)).

The agent respects all the prompt-cache options on CompressionClient (enable_prompt_cache, prompt_cache_ttl, openai_prompt_cache_key). Tavily / Brave keys are read from TAVILY_API_KEY / BRAVE_SEARCH_API_KEY (falls back to BRAVE_API_KEY).

Compression options

Param	Purpose
query	Question the LLM is trying to answer — drives latte_v1 compression
target_compression_ratio	0-1 strength (e.g. 0.5 = remove 50%) or >1 for Nx factor (4 = 4x). Backend max: 200
coarse	True for paragraph-level (default, faster), False for token-level (fine-grained)
heuristic_chunking	Structure-preserving chunking
disable_placeholders	Disable placeholder tokens in output

Error handling

from compresr.exceptions import (
    CompresrError,
    AuthenticationError,
    RateLimitError,
    ValidationError,
)

try:
    result = client.compress(context="...", query="...")
except AuthenticationError:
    print("Invalid API key")
except RateLimitError:
    print("Rate limit exceeded")
except ValidationError as e:
    print(f"Invalid request: {e}")
except CompresrError as e:
    print(f"API error: {e}")

Framework integrations

The agents layer ships in the base install — pip install compresr is enough to get CompressionClient, all three provider chat models (Anthropic / OpenAI / Gemini), and both web search tools (Tavily + Brave).

Genuinely optional integrations beyond the agents layer:

Extra	Pulls in
compresr[langgraph]	langgraph (LangGraph checkpoint serializer, store, handoff tool)
compresr[llamaindex]	llama-index-core (node postprocessor, memory block, tool wrapper)
compresr[litellm]	litellm[proxy] (LiteLLM proxy guardrail)
compresr[all]	all three above

pip install "compresr[langgraph]"

Old compresr[agents] / compresr[agents-anthropic] / compresr[agents-all] / compresr[langchain] install commands still resolve (no-op extras kept for back-compat) — everything they used to pull in is now in the base install.

LangChain — middleware + tool wrapper + retriever

from langchain.agents import create_agent
from compresr.integrations.langchain import (
    CompresrToolMiddleware,
    wrap_tool_with_compression,
    CompresrExtractor,
)

agent = create_agent(
    model=model,
    tools=[web_search],
    middleware=[CompresrToolMiddleware(
        api_key=os.environ["COMPRESR_API_KEY"],
        query_arg="query",
    )],
)

LangGraph — compression as a graph node

from compresr.integrations.langgraph import make_compresr_node

graph.add_node("compress", make_compresr_node(
    api_key=os.environ["COMPRESR_API_KEY"],
    context_key="retrieved_text",
    query_key="user_question",
))

LlamaIndex — node postprocessor for RAG

from compresr.integrations.llamaindex import CompresrNodePostprocessor

query_engine = index.as_query_engine(
    node_postprocessors=[CompresrNodePostprocessor(
        api_key=os.environ["COMPRESR_API_KEY"],
    )],
)

Unified query API

Every integration that accepts a query exposes the same three knobs:

Param	Purpose
query	Static query — same for every call
query_extractor	Callable that derives the query from the call context
query_arg / query_key	Name of the tool arg / state key to use as the query

Priority: query > query_extractor > query_arg/query_key > smart-pick from common arg keys (query, question, search_query, ...) > last human message in history.

Tutorials

Runnable Jupyter notebooks under tutorial/:

• 01_quickstart.ipynb — core CompressionClient.
• 02_langchain.ipynb — middleware + tool wrapper + retriever.
• 03_langgraph.ipynb — compression node in a 3-node graph.
• 04_llamaindex.ipynb — node postprocessor + tool wrapper.
• 05_compresr_agents.ipynb — agent client (Anthropic/OpenAI/native shapes) with auto-compressed tool output.

Requirements

• Python 3.9+
• httpx >= 0.27.0
• pydantic >= 2.10.0
• Optional: langchain>=1.0, langgraph>=0.2, llama-index-core>=0.11 (install the matching extra)

License

Apache 2.0 — see LICENSE.

Support

• Docs: compresr.ai/docs
• Issues: GitHub
• Email: [e-mail]