Appearance
Phase 3: Active Recall Flashcard Hub
Scope: Master Flashcard Decks for Phase 3: Modular Chaining & Complex Workflows
Curriculum Source:docs/plan/03_phase3_modular_chaining_lcel.md
Interaction: Click any card or press Space to flip. Use ← Prev and Next → to navigate.
📚 Deck 1: The Runnable Protocol
Master universal invocation primitives: invoke(), ainvoke(), batch(), and stream():
🗂️ Deck 1: The Runnable Protocol
Card 1 of 3Runnable ProtocolClick or press Space to flip ↺
What are the four core invocation methods guaranteed by any LangChain Runnable?
invoke | ainvoke | batch | streamRunnable Protocol • AnswerClick to flip back ↻
1. invoke(): Synchronous single execution. 2. ainvoke(): Asynchronous non-blocking execution with asyncio. 3. batch(): Parallel concurrent execution of lists with worker pooling. 4. stream(): Real-time token streaming as an iterator.
💡 Architect Takeaway: Any component implementing Runnable can be chained and invoked in any of these 4 modes.
📚 Deck 2: LCEL & Composition Primitives
Master the pipe (|) syntax, RunnableSequence, RunnableParallel, and RunnablePassthrough:
🗂️ Deck 2: LCEL & Composition Primitives
Card 1 of 3LCEL CompositionClick or press Space to flip ↺
How does the Unix pipe (|) operator work in LangChain Expression Language?
chain = prompt | model | parserLCEL Composition • AnswerClick to flip back ↻
The pipe operator composes Runnables into a Directed Acyclic Graph (DAG). The output of the left operand is automatically passed as the input to the right operand, preserving streaming and async capabilities across the entire chain.
💡 Architect Takeaway: LCEL provides a declarative syntax that eliminates imperative orchestration boilerplate.
📚 Deck 3: Provider Interoperability & Standard Content Blocks
Master .content_blocks, unified AIMessage normalization, and multi-cloud model factories:
🗂️ Deck 3: Provider Interoperability
Card 1 of 2Provider InteroperabilityClick or press Space to flip ↺
What architectural problem does the .content_blocks property solve?
Provider Interoperability • AnswerClick to flip back ↻
Different LLM vendors return disparate output schemas (OpenAI choices[0].message vs Anthropic content[0].text vs Gemini parts[0].text). .content_blocks provides a standardized, normalized list of typed blocks (e.g. TextBlock, ToolCallBlock) across all providers.
💡 Architect Takeaway: Interacting with normalized content blocks prevents client code from breaking when swapping models.
📚 Deck 4: Structured Outputs & Schema Adherence
Master .with_structured_output(), JSON schema mode, and immutable Pydantic contracts:
🗂️ Deck 4: Structured Outputs & Schema Adherence
Card 1 of 2Structured OutputsClick or press Space to flip ↺
How does .with_structured_output() guarantee computer-readable JSON over basic prompting?
Structured Outputs • AnswerClick to flip back ↻
Prompting for JSON relies on non-deterministic compliance; models often output markdown wrappers or syntax errors. .with_structured_output(PydanticSchema) configures the provider's native function-calling or JSON schema mode, mathematically constraining token sampling to valid schema tokens and parsing directly into a Pydantic object.
💡 Architect Takeaway: Never parse raw model output strings manually; enforce native structured outputs.
📚 Deck 5: Operational Modifiers (Retries & Fallbacks)
Master .with_retry(), exponential backoff with jitter, .with_fallbacks(), and self-healing pipelines:
🗂️ Deck 5: Operational Modifiers
Card 1 of 2Operational ModifiersClick or press Space to flip ↺
When should you apply .with_retry() versus .with_fallbacks()?
Operational Modifiers • AnswerClick to flip back ↻
Use .with_retry() for transient, recoverable errors (rate limits, network blips, short timeouts) where re-attempting with exponential backoff will likely succeed. Use .with_fallbacks() for persistent failures (provider outage, context window blown, model deprecated) requiring failover to an alternative model.
💡 Architect Takeaway: Combine retries for transient blips with fallbacks for disaster recovery.
📚 Deck 6: Pipeline Instrumentation & Tracing
Master LangFuse nested spans, latency profiling on concurrent branches, and fallback trace evidence:
🗂️ Deck 6: Pipeline Instrumentation
Card 1 of 2Pipeline InstrumentationClick or press Space to flip ↺
How does LangFuse represent a RunnableParallel step in its distributed trace tree?
Pipeline Instrumentation • AnswerClick to flip back ↻
LangFuse captures RunnableParallel as a parent Span with concurrent child spans for each branch. The duration of the parent span reflects the slowest concurrent branch, highlighting the critical path latency bottleneck.
💡 Architect Takeaway: Inspect child span durations to identify which parallel branch is creating latency drag.
📚 Deck 7: Compulsory Security (Moderation & Defensive Schemas)
Master inbound content moderation, PII regex scrubbing, and unauthorized action gating:
🗂️ Deck 7: Compulsory Security
Card 1 of 2Compulsory SecurityClick or press Space to flip ↺
Why should Content Moderation Middleware be placed at the inbound boundary of an LCEL chain?
Compulsory Security • AnswerClick to flip back ↻
Running moderation at the input boundary prevents toxic, hateful, or harmful prompts from consuming downstream LLM tokens, triggering expensive parallel chains, or manipulating internal tools, ensuring early rejection at minimal cost.
💡 Architect Takeaway: Always evaluate safety at the entry boundary before executing expensive model inference.