Tutorial: Getting Started

On this page

2. Create an engine
3. Retrieve documents (external to this library)
4. Build a prompt
5. Send it to your LLM
6. Read back what the engine decided
Next: choose a strategy deliberately

This walks through building a small RAG prompt pipeline from scratch using zerda.prompt.

Python

from zerda.prompt import PromptEngine, Document

2. Create an engine

Python

engine = PromptEngine()

Defaults are production-ready: caching on, auto-detection on, a fresh ContextManager and GuardrailEngine underneath. See Configuration for every override.

3. Retrieve documents (external to this library)

zerda.prompt does not perform retrieval itself — it consumes whatever your retriever returns, in one of three shapes:

Python · 8 lines

# As typed Document objects
docs = [Document(content="...", source="kb/article-1", score=0.91)]

# As plain dicts (looser keys accepted: content/text, source/url, score, metadata)
docs = [{"content": "...", "source": "kb/article-1", "score": 0.91}]

# As plain strings (source/score default)
docs = ["Refunds are processed within 14 business days."]

4. Build a prompt

Python · 4 lines

prompt = engine.build(
    query="How long do refunds take?",
    documents=docs,
)

5. Send it to your LLM

Python · 2 lines

messages = prompt.to_messages()   # OpenAI-shaped
payload = prompt.to_anthropic()   # Anthropic-shaped

6. Read back what the engine decided

Python · 3 lines

print(prompt.strategy)              # PromptStrategy.SIMPLE (default for QA)
print(prompt.guardrails_applied)    # e.g. ['safe_output', 'concise', 'grounding', ...]
print(prompt.estimated_tokens)

Next: choose a strategy deliberately

By default, PromptEngine may auto-detect a strategy from your query (see _auto_detect). For predictable behavior in production, pass strategy= explicitly and consider enable_auto_detect=False at construction time: