src/blu/self_improve.py

import os, json, time
from .rag import build_index, search
from .llm import BluLLM, make_messages

LOG_DIR = ".config/self_improve"
os.makedirs(LOG_DIR, exist_ok=True)

def log_interaction(query: str, response: str, feedback: str=None, metadata: dict=None):
    entry = {
        "timestamp": int(time.time()),
        "query": query,
        "response": response,
        "feedback": feedback or "",
        "meta": metadata or {}
    }
    with open(os.path.join(LOG_DIR, "interactions.jsonl"), "a", encoding="utf-8") as f:
        f.write(json.dumps(entry, ensure_ascii=False) + "\\n")

def generate_synthetic_training(n=50):
    """
    Use Blu to generate high-quality instruction-response pairs from logs.
    These output JSONL suitable for QLoRA SFT.
    """
    out = []
    try:
        lines = open(os.path.join(LOG_DIR, "interactions.jsonl"), "r", encoding="utf-8").read().strip().splitlines()
    except Exception:
        lines = []
    examples = [json.loads(l) for l in lines[-n:]]
    llm = BluLLM()
    for ex in examples:
        prompt = f"Rewrite the user's query into a clear instruction and produce an improved answer with step-by-step reasoning. Query: {ex['query']}. Response: {ex['response']}."
        msgs = make_messages(prompt, system_text="You are an expert data annotator.")
        try:
            improved = llm.chat(msgs)
            out.append({
                "instruction": prompt,
                "input": "",
                "output": improved
            })
        except Exception:
            continue
    out_path = os.path.join(LOG_DIR, "synthetic_for_finetune.jsonl")
    with open(out_path, "w", encoding="utf-8") as f:
        for o in out:
            f.write(json.dumps(o, ensure_ascii=False) + "\\n")
    return out_path