Cookbook: AI Recipes & Patterns

Short, focused recipes that solve specific problems. Unlike tutorials, which build understanding progressively, cookbook recipes are self-contained solutions you can apply directly. Each recipe includes the problem statement, solution code, and an explanation of why the approach works. Browse by category or search for a specific topic.

Recipe Categories

Category	Recipes	Description
Agent Recipes	Agent tool failures, parallel execution, custom patterns. Handle real-world agent challenges: graceful tool failure recovery, concurrent tool execution for throughput, and reusable agent patterns that compose with the BaseAgent abstraction.
LLM Recipes	History trimming, streaming metadata, token counting, error handling. Solve common LLM integration problems: manage conversation history within token limits, extract metadata from streaming responses, count tokens accurately across providers, and implement resilient error handling with retry strategies.
RAG Recipes	Batch embeddings, parent document retrieval, Cohere reranking. Optimize retrieval pipelines: batch embedding operations for throughput, parent document retrieval for preserving context around matched chunks, metadata-based filtering, and reranking with cross-encoder models for precision.
Memory Recipes	TTL cleanup, context recovery, conversation expiry. Manage agent memory at scale: configure time-to-live policies for automatic cleanup, recover conversation context after service restarts, and implement expiry policies that balance storage costs with context availability.
Voice Recipes	Stream scaling, backends, speech interruption, latency optimization. Tune production voice pipelines: scale concurrent audio streams, configure backend providers for your deployment, handle speech interruptions naturally, measure and optimize glass-to-glass latency, and implement jitter buffering for network resilience.
Multimodal Recipes	Multiple images, inbound media handling. Process diverse content types: batch-process multiple images in a single LLM call, handle inbound media from messaging platforms with automatic format detection, and implement capability-based fallbacks when providers lack multimodal support.
Prompt Recipes	Dynamic templates, partial substitution. Build flexible prompt systems: construct prompts dynamically from runtime data, use partial template substitution for reusable prompt fragments, and manage prompt versioning across environments.
Infrastructure Recipes	Config hot reload, rate limiting, request ID correlation. Solve cross-cutting production concerns: hot-reload configuration without restarting, enforce per-tenant rate limits, propagate request IDs through distributed traces, detect and redact PII in LLM payloads, validate schemas at system boundaries, and manage circuit breakers for external dependencies.

Each category contains a comprehensive overview plus individual recipe pages. Browse the sidebar categories to find specific topics.

Recipe Format

Each recipe follows a consistent pattern:

1. Problem — What you are trying to solve and why it matters in production
2. Solution — Working code you can copy directly into your project
3. Explanation — Why the approach works, what tradeoffs it makes, and when to consider alternatives