chain-llm-pattern

Chain LLM Pattern for n8n

Multi-step LLM chains outperform single-shot prompts on any task that combines extraction + reasoning. This skill encodes the production pattern.

When to use a chain vs a single prompt

| Single prompt works | Chain is better | |---|---| | "Summarize this email" | "Extract entities, then categorize by urgency, then decide routing" | | "Translate this to English" | "Detect language, translate, then extract structured fields" | | "Is this spam? yes/no" | "Score spam probability from email, phone, IP, content separately, then combine" |

Rule of thumb: if the task has ≥2 distinct reasoning steps OR the final decision depends on intermediate structured data, use a chain.

The pattern (3-stage default)

Input → [Extract] → [Analyze/Classify] → [Score/Decide] → Output

Each stage is its own LLM node with its own prompt. Between stages, use Set or Code nodes to transform and validate.

Stage 1: Extract (schema-bound)

Use Information Extractor node (LangChain). NOT a generic AI Agent or raw HTTP call.

Why: Information Extractor binds output to a JSON schema. It parses, retries on invalid JSON, and fails loudly — instead of silently returning prose you then regex.

Define schema explicitly:

{
  "type": "object",
  "properties": {
    "customer_name": { "type": "string" },
    "product_mentioned": { "type": "string" },
    "sentiment": { "enum": ["positive", "neutral", "negative"] },
    "urgency_score": { "type": "number", "minimum": 0, "maximum": 10 }
  },
  "required": ["customer_name", "sentiment"]
}

System prompt for this stage: short, one job. "Extract the fields defined in the schema from the transcript. If a field is absent, omit it. Do not infer or guess."

Stage 2: Analyze (reason over extracted data)

Use Basic LLM Chain with the extracted JSON from Stage 1 as input.

This stage reasons: categorize, cluster, identify patterns, detect issues. The input is structured (from Stage 1) so the model isn't juggling parsing + reasoning simultaneously.

Example system prompt:

Given the extracted customer data below, classify into one of: [technical_issue, billing_question, cancellation_risk, upsell_opportunity]. Then identify the single most important next action. Return JSON with category and next_action.

Stage 3: Score / decide (deterministic where possible)

If the final step is arithmetic (e.g., composite scoring: 0.4 × email_score + 0.3 × phone_score + 0.3 × content_score), use a Code node, NOT an LLM.

LLMs are bad at arithmetic. They fail silently. Use Code (JavaScript) for any math involving weights, thresholds, or aggregation.

Model selection

| Stage | Recommended model | Why | |---|---|---| | Extract | Groq llama-3.3-70b-versatile or openai/gpt-4o-mini | Fast, cheap, good at schema adherence | | Analyze | Claude Sonnet 4 or GPT-4o | Reasoning quality matters more | | Score (if LLM) | gpt-4o-mini | Arithmetic weakness, keep cheap |

Groq is the fastest provider for extract stages — 500+ tokens/sec. Use it unless you need Claude/OpenAI specifically.

Production rules

1. Pin example data at each stage during development. Right-click node → "Pin Data". Without pinning, changing Stage 1 invalidates all downstream test data and you waste API calls.

1. Budget tokens explicitly. Set maxTokens on every LLM node. Stage 1 extract rarely needs >500. Stage 2 analyze rarely >1000.

1. Validate between stages. Insert a Code node between LLM stages that checks required fields exist. Fail fast with a clear error — don't let a missing field propagate and produce a confusing Stage 3 failure.

1. Log stage outputs. Add a MySQL or Google Sheets insert after Stage 1 and Stage 2 that records the raw output (truncated to 1000 chars). You WILL need this for debugging.

1. Temperature: 0 for extract, 0.2–0.4 for analyze, 0 for scoring. Extract must be deterministic. Analysis benefits from slight variance. Scoring must be deterministic.

Common multi-language variant (EN/ES/PT)

For transcripts in mixed languages, add a Stage 0:

Stage 0 (Groq): Detect language → route to language-specific prompts
Stage 1 (language-specific): Extract in source language
Stage 2: Translate structured output to English (cheap, short)
Stage 3: Analyze in English

Language-specific prompts extract better than a single multilingual prompt because entity names (cities, products) follow different patterns per language.

Anti-patterns

• Don't use AI Agent for extraction. Agents are for tool use, not structured output. Use Information Extractor.
• Don't concatenate all steps into one mega-prompt. Each hallucination compounds. Separate stages let you evaluate each independently.
• Don't loop an LLM on retry without a counter. Infinite loops cost money. Cap retries at 3 via a counter in a Set node.

Reference

• references/groq-chain-example.json — a working 4-node chain ready to import into n8n