RLEngine API

RLEngine extends InferenceEngine with self-improvement strategies and SQLite memory.

Class Definition

class RLEngine(InferenceEngine):
    def __init__(
        self,
        model_name: str,
        device: Optional[str] = None,
        bits: int = 16,
        offload: bool = False,
        memory_db: str = None,
    )

Methods

generate()

def generate(
    self,
    prompt: str,
    max_tokens: int = 128,
    strategy: str = "auto"
) -> str

Generate with RL self-improvement and memory.

Parameters:

Parameter	Type	Default	Description
prompt	str	required	Input prompt
max_tokens	int	128	Maximum tokens to generate
strategy	str	"auto"	Strategy: "auto", "best_of_n", "refine", "expand", "consensus"

Returns: Generated text string.

get_stats()

def get_stats(self) -> Dict

Get generation statistics.

Returns:

{
    "generations": 100,
    "memory_hits": 25
}

search_memory()

def search_memory(
    self,
    prompt: str,
    limit: int = 3
) -> List[Dict]

Search memory for similar prompts.

Parameters:

Parameter	Type	Default	Description
prompt	str	required	Search query
limit	int	3	Maximum results

Returns:

[
    {
        "prompt": "...",
        "response": "...",
        "strategy": "refine",
        "quality_score": 0.8
    }
]

Inner Classes

QualityScorer

class QualityScorer:
    def score(self, text: str) -> float

Score a response (0.0 - 1.0).

MemoryDB

class MemoryDB:
    def __init__(self, db_path: str = "llmforge_memory.db")
    def store(self, prompt: str, response: str, strategy: str, quality_score: float)
    def recall(self, prompt: str, limit: int = 3) -> List[Dict]
    def get_best(self, prompt: str) -> Optional[str]
    def close(self)

SQLite-based memory storage.