ReLLA — Reversible LLM Adapter

ReLLA is a lightweight, offline‑first, provider‑agnostic LLM adapter designed for developers who need reliability, reproducibility, and control over LLM execution.

It records every LLM call as a deterministic artifact in SQLite, enabling replay, caching, migration, and benchmarking across providers and frameworks — without rewriting application code.

ReLLA is intentionally minimal. It is infrastructure tooling, not a framework.

---

Core Capabilities

• Deterministic request hashing and artifact storage
• SQLite‑backed replay and cache layer
• Provider‑agnostic execution (local and cloud)
• Framework‑independent core engine
• Fully offline‑testable and reproducible
• Designed to run on low‑resource machines

---

Installation

Core Engine

pip install reversible-lla

This installs the ReLLA core only. No framework dependencies are included by default.

Optional Framework Bridges

Framework integrations are opt‑in and installed explicitly:

pip install reversible-lla[langchain]
pip install reversible-lla[llamaindex]
pip install reversible-lla[langchain,llamaindex]

The core engine remains unchanged regardless of framework usage.

---

Command Line Interface

The CLI provides a direct, human‑friendly interface to the ReLLA engine.

Execute and Record

rella run "hello world" --db rella.db

This executes the request and stores the result as an immutable artifact.

Replay from Cache

rella replay "hello world" --db rella.db

If a matching artifact exists, the stored output is returned without invoking any provider.

---

Artifact Storage

All executions are stored in a local SQLite database.

sqlite3 rella.db
.tables
SELECT * FROM artifacts;

Each artifact contains:

• Deterministic input hash
• Provider identifier
• Model identifier
• Output payload

---

LangChain Integration

ReLLA can be used as a drop‑in LLM within LangChain applications.

from rella.bridges.langchain import ReLLALLM

llm = ReLLALLM(provider="mock", db="rella.db")
print(llm.invoke("hello from langchain"))

Artifacts generated via LangChain are immediately available for replay via the CLI.

---

LlamaIndex Integration

ReLLA integrates with LlamaIndex by implementing the full LLM interface.

from rella.bridges.llamaindex import ReLLALLM

llm = ReLLALLM(provider="mock", db="rella.db")
response = llm.complete("hello from llamaindex")
print(response.text)

Execution and replay semantics remain identical across all entry points.

---

Design Principles

• Reproducibility over convenience
• Determinism over heuristics
• Minimal surface area
• Explicit dependency control
• Offline‑first execution

---

Non‑Goals

• Agent orchestration
• Prompt chaining frameworks
• Workflow automation
• Hosted services or dashboards

---

Release Status

v0.1.0 — Core engine and framework bridges are stable.

---

License

MIT