contextcram

Fit anything into an LLM context window. A tiny, zero-dependency, priority-aware token-budget packer for RAG pipelines and agents.

Every RAG or agent app has the same problem: you have too much stuff — a system prompt, chat history, retrieved documents, tool output — and a fixed token budget. contextcram packs it all in by priority, truncating, trimming, or dropping the least important pieces so the important ones always make it.

from contextcram import Packer

packer = Packer(budget=8000)  # token budget

packer.add(system_prompt, priority="required")                 # never dropped
packer.add(chat_history, priority="high", strategy="trim")     # drop oldest turns
packer.add(retrieved_docs, priority="medium", strategy="drop") # all-or-nothing
packer.add(tool_output, priority="low", strategy="truncate")   # cut to fit

result = packer.fit()
print(result.text)            # assembled, in-budget context
print(result.used_tokens)     # e.g. 7840
print(result.dropped_names)   # what didn't make the cut

Why

• Zero dependencies. Pure Python. Works out of the box with a fast characters-per-token heuristic; plug in tiktoken or any tokenizer when you need exact counts.
• Framework-agnostic. Use it with LangChain, LlamaIndex, the raw provider SDKs, or nothing at all.
• Priority-aware. You decide what survives a tight budget, not a blind truncate at the end.
• Observable. Every result tells you what was kept, truncated, and dropped.

Installation

pip install contextcram
# optional: exact token counts via tiktoken
pip install "contextcram[tiktoken]"

Strategies

When an optional item doesn't fully fit, its strategy decides what happens:

required items are always kept; if they alone exceed the budget, a BudgetExceeded error is raised.

Model-aware budgets

Skip the magic number — set the budget from the model, and reserve room for the response in one go:

from contextcram import Packer

# 128k window for gpt-4o, holding back 2k tokens for the model's reply
packer = Packer(model="gpt-4o", reserve=2000)
print(packer.full_budget)  # 128000
print(packer.budget)       # 126000  (effective budget you pack into)

reserve is the easy way to avoid the classic "prompt fit, but no room left to answer" failure. Unknown model? Pass budget= explicitly or register it:

from contextcram import register_model

register_model("my-internal-llm", 32000)
packer = Packer(model="my-internal-llm", reserve=1000)

Exact token counts

from contextcram import Packer, tiktoken_tokenizer

packer = Packer(budget=8000, tokenizer=tiktoken_tokenizer("gpt-4o"))

Or wrap any tokenizer with CallableTokenizer(lambda s: len(my_encode(s))).

Priorities

Use the named levels "required", "high", "medium", "low", or pass any integer (higher is kept first):

packer.add(text, priority=42, strategy="truncate")

Real-world usage

With LangChain

Pack a system prompt, retrieved docs, and chat history into a gpt-4o budget — leaving room for the answer — then hand the result to the model:

from contextcram import Packer
from langchain_openai import ChatOpenAI
from langchain_core.messages import SystemMessage, HumanMessage

llm = ChatOpenAI(model="gpt-4o")

docs = [d.page_content for d in retriever.invoke(question)]
history = [f"{m.type}: {m.content}" for m in memory.messages]

ctx = (
    Packer(model="gpt-4o", reserve=1500)                          # room for the reply
    .add(SYSTEM_PROMPT, priority="required")                      # never dropped
    .add(history, priority="high", strategy="trim")               # drop oldest turns
    .add("\n\n".join(docs), priority="medium", strategy="drop")   # whole docs only
    .fit()
)

response = llm.invoke([SystemMessage(ctx.text), HumanMessage(question)])

With the raw Anthropic SDK

Tie reserve to max_tokens so the input can never crowd out the response:

import anthropic
from contextcram import Packer

client = anthropic.Anthropic()
REPLY_TOKENS = 4000

ctx = (
    Packer(model="claude-opus-4-8", reserve=REPLY_TOKENS)
    .add(SYSTEM_PROMPT, priority="required")
    .add(chat_history, priority="high", strategy="trim")
    .add(retrieved_docs, priority="medium", strategy="drop")
    .fit()
)

msg = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=REPLY_TOKENS,        # matches reserve above
    system=ctx.text,
    messages=[{"role": "user", "content": question}],
)
print(f"packed {ctx.used_tokens} tokens; dropped {ctx.dropped_names}")

Alternatives

Priority-based context assembly isn't a new idea, and depending on your needs one of these may fit better — contextcram deliberately trades features for simplicity and zero dependencies:

Choose contextcram when you want a tiny, zero-dependency, framework-agnostic helper with a 3-line API (Packer(...).add(...).fit()) that does one thing — fit prioritized pieces into a budget — and gets out of your way.

Development

git clone https://github.com/Waelr1985/contextcram.git
cd contextcram
uv sync
uv run pytest
uv run ruff check .
uv run mypy

License

MIT