rag-cachex 0.1.1

A context-aware, multi-level caching system for Retrieval-Augmented Generation (RAG) applications.

RAGCache

RAGCache is a context-aware, multi-level caching system designed to accelerate and optimize Retrieval-Augmented Generation (RAG) pipelines. Unlike traditional exact-match KV caches, RAGCache utilizes semantic similarity, context stability validation, and user intent classification to safely reuse LLM responses while strictly preserving correctness.

---

🚀 Key Features

• Dual-Layer Caching (L1 & L2):
  • L1 (Retrieval Cache): Low-latency exact-match string cache mapping queries directly to document IDs, bypassing heavy vector search lookups.
  • L2 (Generation Cache): High-performance semantic cache mapping queries and contexts to LLM responses.
• Context Stability Validation: Analyzes document overlaps using Jaccard Similarity to reject cache hits if retrieved context documents have drifted.
• Intent-Aware Caching: Classifies queries (e.g. action, informational) to bypass caching automatically for mutation operations.
• Strict Tenant Isolation: Restricts cache retrieval scope dynamically to prevent context leakages across different users and tenants.
• Prometheus Observability: Native latency histograms, cache hit-rate counters, and memory tracking.
• Optimization ROI: Speeds up RAG pipelines by 2.3x and reduces API token costs by up to 60% with zero loss in contextual accuracy.

---

📐 Architecture Overview

                        User Query
                            │
                            ▼
                    ┌──────────────┐
                    │  L1 Cache    ├─────────► Hit (Yields Doc IDs)
                    └──────┬───────┘                 │
                           │ Miss                    │
                           ▼                         │
                    ┌──────────────┐                 │
                    │ Vector DB    │◄────────────────┘
                    └──────┬───────┘
                           │ doc_ids
                           ▼
                    ┌──────────────┐
                    │  L2 Cache    ├─────────► Hit (Returns LLM Response)
                    └──────┬───────┘
                           │ Miss
                           ▼
                    ┌──────────────┐
                    │  Expensive   │
                    │   LLM API    │
                    └──────────────┘

For a deep dive into caching layers and algorithms, read the Architecture Guide.

---

📦 Installation

Install the core library (includes Redis support and Prometheus telemetry):

pip install rag-cache

Optional Backends:

• Local FAISS Vector DB:

  pip install "rag-cache[faiss]"
  

• Local Sentence-Transformers Embeddings:

  pip install "rag-cache[embeddings]"
  

• Full Installation (All Backends):

  pip install "rag-cache[all]"
  

---

⚡ Quick Start

Get running in less than 2 minutes using our high-level facade:

from rag_cache import RAGCache

# Initialize dual-layer cache facade
cache = RAGCache()

# Wrap your existing RAG execution function
def query_rag_pipeline(query: str):
    # Pass query, retriever callback, and LLM callback
    result = cache.run(
        query=query,
        retriever=lambda q: ["doc_abc", "doc_xyz"],
        llm=lambda q, doc_ids: "Context-aware systems ensure LLM correctness."
    )
    return result["answer"]

# Executes LLM call (miss)
print(query_rag_pipeline("What is context-aware caching?"))

# Returns cached response immediately (semantic hit!)
print(query_rag_pipeline("Explain context-aware caching."))

For more execution models and advanced configurations, see the Quick Start Guide.

---

⚙️ Configuration

RAGCache supports configuration loading with the following precedence:

1. Keyword overrides in constructor (e.g. RAGCache(redis_url="...")).
2. Environment variables (REDIS_URL, L1_TTL, L2_TTL, SIMILARITY_THRESHOLD).
3. YAML configuration file (e.g. loaded via RAGCache.from_config("config.yaml")).
4. Built-in defaults.

See config_example.yaml for a complete template configuration.

---

🔒 Tenant Isolation

Ensure user data boundaries are strictly isolated by passing a tenant_id to caching endpoints:

result = cache.run(
    query=query,
    retriever=retriever_fn,
    llm=llm_fn,
    tenant_id="customer_company_a",
    scope="tenant"  # Options: "tenant" (default), "user", or "global"
)

For details on namespacing strategies in Redis, see Architecture Guide.

---

📊 Metrics & Telemetry

RAGCache collects statistics out-of-the-box using Prometheus:

• Cache hits/misses split by L1 and L2 layers.
• Operational latency histograms for Redis, FAISS, and Embeddings.
• Eviction count and Redis memory usage.

To enable the HTTP telemetry server, set the prometheus_port config parameter:

cache = RAGCache(prometheus_port=8000)

Scrape metrics at http://localhost:8000/metrics. We provision pre-built Grafana dashboards under config/grafana/dashboards.

---

🧪 Calibration & Benchmarking

To optimize caching threshold parameters for your specific RAG datasets, run the offline calibration script:

python tools/calibration/calibrate.py

This runs a threshold sweep to recommend optimal parameters under Safety-First, Max F1, and Balanced profiles. Read the Benchmark Documentation for more details.

---

📄 License

This project is licensed under the MIT License. See LICENSE for details.