LoopyLuci Advanced Few-Shot Prompting
Implementing advanced few-shot prompting techniques — from dynamic example selection through chain-of-thought, structured outputs, and multi-turn example management.
When to Use
- Improving LLM output quality with minimal examples
- Dynamic example selection based on query similarity
- Few-shot chain-of-thought for reasoning tasks
- Managing context window with example compression
Few-Shot Methods
FEW_SHOT_METHODS = {
'static': 'Fixed examples in every prompt — simple, but may not match query',
'dynamic_knn': 'Retrieve examples from database via embedding similarity — best match',
'auto_generated': 'LLM generates its own examples for the task — no labeled data needed',
'clustered': 'Select diverse examples covering different task types/classes',
'compressed': 'Summarize examples to fit more in context — trade detail for quantity',
}
class DynamicFewShot:
"""Dynamic example selection using embedding similarity."""
def __init__(self, examples: List[Dict], encoder):
self.examples = examples
self.encoder = encoder
self.example_embeddings = encoder.encode([e['query'] for e in examples])
def select_examples(self, query: str, k: int = 3) -> List[Dict]:
from sklearn.metrics.pairwise import cosine_similarity
query_emb = self.encoder.encode([query])
similarities = cosine_similarity(query_emb, self.example_embeddings)[0]
top_k = similarities.argsort()[-k:][::-1]
return [self.examples[i] for i in top_k]
Verification Checklist
- Few-shot method chosen (static, dynamic, auto-generated)
- Examples representative of expected queries
- Dynamic selection performance (embedding similarity latency)
- Context window budgeted (examples + instructions + query fit in context)
- Example diversity maintained (not all similar examples)
- Accuracy compared with zero-shot baseline