LangChainLangChain避坑指南

LCEL Basics

• - | pipes output to next — INLINECODE1
• INLINECODE2 forwards input unchanged — use in parallel branches
• INLINECODE3 runs branches concurrently — INLINECODE4
• INLINECODE5 for single, .batch() for multiple, .stream() for tokens
• Input must match expected keys — {"question": x} not just x if prompt expects INLINECODE10

Memory Gotchas

• - Memory doesn't auto-persist between sessions — save/load explicitly
• INLINECODE11 grows unbounded — use ConversationSummaryMemory for long chats
• Memory key must match prompt variable — memory_key="chat_history" needs {chat_history} in prompt
• INLINECODE15 for chat models — False returns string for completion models

RAG Chunking

• - Chunk size affects retrieval quality — too small loses context, too large dilutes relevance
• Chunk overlap prevents cutting mid-sentence — 10-20% overlap typical
• INLINECODE17 preserves structure — splits on paragraphs, then sentences
• Embedding dimension must match vector store — mixing models causes silent failures

Output Parsers

• - PydanticOutputParser needs format instructions in prompt — call INLINECODE19
• Parser failures aren't always loud — malformed JSON may partially parse
• INLINECODE20 retries with LLM — wraps another parser, fixes errors
• INLINECODE21 on chat models — cleaner than manual parsing for supported models

Retrieval

• - similarity_search returns documents — .page_content for text
• INLINECODE24 parameter controls results count — more isn't always better, noise increases
• Metadata filtering before similarity — filter={"source": "docs"} in most vector stores
• INLINECODE26 for diversity — avoids redundant similar chunks

Agents

• - Agents decide tool order dynamically — chains are fixed sequence
• Tool descriptions matter — agent uses them to decide when to call
• INLINECODE27 — prevents crash on malformed agent output
• Max iterations prevents infinite loops — max_iterations=10 default may be too low

Common Mistakes

• - Prompt template variables case-sensitive — {Question} ≠ INLINECODE30
• Chat models need message format — ChatPromptTemplate, not INLINECODE32
• Callbacks not propagating — pass config={"callbacks": [...]} through chain
• Rate limits crash silently sometimes — wrap in retry logic
• Token count exceeds context — use trim_messages or summarization for long histories