llm-app-patterns

LLM Application & AI Assistant Patterns

Resources

Architecture decision matrix

Pattern	Use when	Cost
Simple RAG	FAQ, docs Q&A	Low
Hybrid RAG (semantic + BM25)	Mixed query types	Medium
Function calling	Structured tool use	Low
ReAct agent	Multi-step reasoning	Medium
Plan-and-execute	Complex decomposable tasks	High
Multi-agent	Research, critique-refine	Very High

RAG: critical config numbers

CHUNK_CONFIG = {
    "chunk_size": 512,       # tokens — sweet spot for most docs
    "chunk_overlap": 50,     # prevents context loss at boundaries
    "separators": ["\n\n", "\n", ". ", " "],
}
# Hybrid search alpha: 1.0=semantic only, 0.0=BM25 only, 0.5=balanced

RAG: retrieval strategies

# Basic: semantic search
results = vector_db.similarity_search(embed(query), top_k=5)

# Better: hybrid (semantic + keyword via RRF)
def hybrid_search(query, alpha=0.5):
    return rrf_merge(vector_db.search(query), bm25_search(query), alpha)

# Best for recall: multi-query (3 variations, deduplicate)
queries = llm.generate_variations(query, n=3)
results = deduplicate([semantic_search(q) for q in queries])

RAG: generation prompt template

RAG_PROMPT = """Answer based ONLY on the context below.
If insufficient, say "I don't have enough information."

Context: {context}
Question: {question}
Answer:"""

Agent: function calling loop

messages = [{"role": "user", "content": question}]
while True:
    response = llm.chat(messages=messages, tools=TOOLS, tool_choice="auto")
    if not response.tool_calls:
        return response.content
    for call in response.tool_calls:
        result = execute_tool(call.name, call.arguments)
        messages.append({"role": "tool", "tool_call_id": call.id, "content": str(result)})

Production: caching (only temperature=0 responses)

def get_or_generate(prompt, model, **kwargs):
    deterministic = kwargs.get("temperature", 1.0) == 0
    if deterministic:
        key = sha256(f"{model}:{prompt}:{json.dumps(kwargs, sort_keys=True)}")
        if cached := redis.get(key): return cached
    response = llm.generate(prompt, model=model, **kwargs)
    if deterministic: redis.setex(key, 3600, response)
    return response