layer-context 0.1.5

Context Layer SDK - versioned context infrastructure for AI apps

Context Layer

Python SDK for ContextLayer - versioned context infrastructure for AI apps.

Installation

pip install layer-context

For client-side encryption (optional):

pip install layer-context cryptography

Quick Start

from layer_context import CL

cl = CL(api_key="cl_sk_your_key_here")

# Create a block
cl.create_block("system-prompt")

# Get a block handle
block = cl.block("system-prompt")

# Append context
block.append("You are a helpful assistant.")
block.append("Always respond in markdown.")

# Get compiled context
print(block.get_context())
# Output: "You are a helpful assistant.\nAlways respond in markdown."

# Update (replace entire context)
block.update("You are a coding assistant. Use Python examples.")

# Delete specific content
block.delete("Use Python examples.")

# Read updated context
print(block.get_context())
# Output: "You are a coding assistant."

# Create a new version
block.create_version()

# Work with specific versions
block.append("V2 content", version=2)
print(block.get_context(version=2))

# View history
entries = block.get_history()

# List all blocks
blocks = cl.list_blocks()

Client-Side Encryption

Enable zero-knowledge encryption so your context is encrypted before it leaves your machine:

cl = CL(
    api_key="cl_sk_your_key_here",
    encryption_key="my-secret-passphrase"
)

block = cl.block("private-data")
block.append("Sensitive context here")

# Data is stored encrypted on the server
# Only you can read it with the same encryption_key
print(block.get_context())  # Decrypted automatically

Note: Requires pip install cryptography. The server never sees your plaintext data.

Context Compression

Compress your context locally using sentence embeddings, TextRank, and TF-IDF — no LLM or API keys needed.

Installation

pip install layer-context[compression]

This installs sentence-transformers and numpy. The all-MiniLM-L6-v2 model (~80MB) is downloaded once on first use.

Usage

block = cl.block("support_faq")
block.append("You are a customer support assistant for Acme Corp.")
block.append("Always greet the user warmly and ask how you can help.")
block.append("If the user asks about billing, explain our pricing tiers.")
block.append("We offer Basic ($9/mo), Pro ($29/mo), and Enterprise (custom).")
block.append("For password resets, direct users to settings > security.")
block.append("Generally speaking, you should always be polite and helpful.")
block.append("Our refund policy allows refunds within 30 days of purchase.")

compressed = block.get_compressed_context(
    query="How much does the Pro plan cost?",
    target_tokens=50,
    aggressive_pruning=True,
)

print(compressed)
# Output (approx): "We offer Basic ($9/mo), Pro ($29/mo), and Enterprise (custom).
#   If the user asks about billing, explain our pricing tiers."
# Original: ~80 tokens -> Compressed: ~30 tokens

Arguments

Argument	Type	Default	Description
query	str	required	Topic/question to optimize relevance for. Sentences irrelevant to this are dropped.
version	int	1	Which block version to compress.
target_tokens	int | None	None	Max token budget for output. Defaults to ~50% of original token count.
relevance_threshold	float	0.3	Min cosine similarity to query to keep a sentence. Lower = keep more, higher = stricter.
dedup_threshold	float	0.85	Similarity above which sentences are deduplicated. Only the best representative is kept.
aggressive_pruning	bool	False	When True, strips filler words, hedge phrases, verbose constructions, and example clauses.

Convenience Shortcut

# Equivalent to cl.block("support_faq").get_compressed_context(...)
compressed = cl.get_compressed("support_faq", query="billing question")

How It Works

Raw context
  → Sentence splitting
  → Relevance filter (cosine similarity vs query ≥ threshold)
  → Semantic dedup (cluster similar sentences, keep best)
  → TextRank + TF-IDF density scoring
  → Token budget selection
  → Optional aggressive pruning
  → Reorder by original position
Compressed context

The entire pipeline runs locally on your machine. No API calls, no extra costs.

License

MIT