smart-cache-package 0.1.3

An intelligent caching and LLM routing package using Mem0AI and zero-shot classification.

Smart Cache Package

Smart Cache Package is a pip-installable Python library that provides an intelligent caching and LLM routing solution. It efficiently reduces redundant LLM calls while ensuring that queries benefit from relevant past interactions.

This package leverages:

• Mem0AI for storing and retrieving past interactions.
• Hugging Face's zero-shot classification to automatically assign categories to queries.
• Built-in LLM integrations for OpenAI (default), Anthropic (coming soon), and support for custom LLM callables.

---

🚀 Features

✅ Smart Caching and Query Reuse

• Caches previous query–answer pairs to avoid unnecessary LLM calls.
• Uses Mem0AI to store user interactions for future reference.

🔍 Near-Duplicate Detection

• Searches Mem0 for semantically similar queries.
• If a near-duplicate query is found (above a configurable similarity threshold), it retrieves the stored response.

📄 Context Building for Richer Responses

• If no exact or near-duplicate answer exists, it retrieves related stored interactions to provide context to the LLM.
• Ensures the LLM receives additional context, improving coherence without exceeding token limits.

🔄 Built-in LLM Routing

• Supports multiple LLM providers:
  • OpenAI (Default): Uses OPENAI_API_KEY from environment variables.
  • Custom LLM: Users can provide their own callable function.

🔄 Export & Import Cache Data (New Feature)

• Users can export their cache to a file and import it into another instance of SmartCache.
• This feature enables **collaboration and

🛠 Debugging and Transparency

• get_answer() can return an extra debug flag indicating the source of the answer:
  • "Local Cache" → Retrieved from in-memory cache.
  • "Mem0 Near-Duplicate" → Retrieved from Mem0.
  • "LLM Call" → Fetched from an external LLM.

---

📥 Installation

pip install smart_cache_package

PREREQUISITES

• Make sure you export OPENAI_API_KEY even if you use a custom LLM, it is required for storing the embeddings using the embedding model from openai.
• Custom Embedding Model support coming soon.

export OPENAI_API_KEY = <your_openai_key>

🚀 Usage

1️⃣ Initialize SmartCache

The package automatically picks OpenAI as the default LLM if no custom callable is provided.

from smart_mem_cache import SmartCache

# Initialize the cache system (requires OPENAI_API_KEY)
smart_cache = SmartCache(
    similarity_threshold_reuse=0.75,
    similarity_threshold_context=0.4,
    max_context_tokens=256,
    ttl_seconds=60,  # 1-minute TTL
    debug=True,
    llm_name="openai"  # Use OpenAI by default
)

user_id = "alice"

Store Initial Interactions

Store a user’s conversation history for retrieval later.

smart_cache.store_interaction_auto_cat(
    user_id, 
    "Hi, I'm Alex. I like to play cricket on weekends.", 
    "Hello Alex! Great to know you enjoy cricket. I'll keep that in mind."
)

smart_cache.store_interaction_auto_cat(
    user_id, 
    "I also like to workout and cook on the weekends", 
    "That's wonderful to have hobbies."
)

print("Stored initial interactions.")

Ask a New Question

If an exact match or near-duplicate exists, it will reuse the previous answer. Otherwise, it calls the LLM.

new_question = "What does Alex do for fun on Saturday?"
answer, source = smart_cache.get_answer(user_id, new_question, return_debug=True)

print(f"Question: {new_question}")
print(f"Answer: {answer}")
print(f"Source: {source}")  # Could be "Local Cache", "Mem0 Near-Duplicate", or "LLM Call"

Ask an Exact Match Question

If a user repeats a previously asked query, it should return from cache.

exact_question = "Does Alex like to play cricket on weekends?"

# First call (expected LLM call)
answer1, source1 = smart_cache.get_answer(user_id, exact_question, return_debug=True)
print(f"First Call - Source: {source1}")

# Second call (should return from cache)
answer2, source2 = smart_cache.get_answer(user_id, exact_question, return_debug=True)
print(f"Second Call - Source: {source2}")  # Expected: "Local Cache"

Force Refresh to Bypass Cache

If you need to force a fresh answer from LLM, use force_refresh=True.

forced_question = "Any weekend hobbies for Alex?"
answer, source = smart_cache.get_answer(user_id, forced_question, force_refresh=True, return_debug=True)

print(f"Forced Refresh - Source: {source}")  # Expected: "LLM Call"

Provide Negative Feedback (Remove Cached Answer)

smart_cache.user_feedback(user_id, exact_question, helpful=False)
print("Negative feedback provided, answer removed from cache.")

# Asking again should now trigger a fresh LLM call.
answer, source = smart_cache.get_answer(user_id, exact_question, return_debug=True)
print(f"Re-asked after negative feedback - Source: {source}")  # Expected: "LLM Call"

Export & Import Cached Data (New Feature)

Exporting Cache to a JSON File:

smart_cache.export_cache("cache_data.json")
print("Cache exported successfully.")

Importing Cache from a JSON File:

smart_cache.import_cache("cache_data.json")
print("Cache imported successfully.")

Using a Custom LLM (Instead of OpenAI)

If you want to use a custom LLM (e.g., a self-hosted model), provide a callable function.

def custom_llm(prompt: str) -> str:
    return f"[Custom LLM response for: {prompt[:50]}...]"

# Initialize SmartCache with custom LLM
smart_cache = SmartCache(
    similarity_threshold_reuse=0.75,
    similarity_threshold_context=0.4,
    max_context_tokens=256,
    ttl_seconds=60,
    debug=True,
    llm_callable=custom_llm  # Pass custom LLM function
)

Environment Variables (Required for OpenAI & Anthropic)

If you use OpenAI or Anthropic, ensure the API keys are set as environment variables.

export OPENAI_API_KEY="your-openai-api-key"
export ANTHROPIC_API_KEY="your-anthropic-api-key"

🚀 Enjoy smart caching and optimized LLM interactions! 🚀