Dify workflow patterns — 10 we actually use in production

Pattern 1 — KB fallback#

Problem: top-1 retrieval below threshold shouldn’t be answered confidently.

Do: “Retrieval → conditional branch” — if score < 0.5, reply “No relevant info, handing off.”

Pattern 2 — Context compression#

Problem: long history blows up tokens.

Do: “Get context → LLM summarize → replace history with summary.” Compression ratio typically 5-10×.

Pattern 3 — Intent + routing#

Problem: pre-sales / post-sales / complaints in one bot performs poorly.

Do: front-load an LLM classifier (enum output), branch on it.

Pattern 4 — Concurrent tool calls#

Problem: looking up order + shipping + coupons serially is slow.

Do: “Iterate” or “parallel” nodes — fire three HTTP calls, merge results.

Pattern 5 — Retry on failure#

Problem: HTTP tools occasionally 500.

Do: HTTP node “3 retries with exponential backoff” + fallback branch returns a degraded answer.

Pattern 6 — Rate limiting#

Problem: promo day floods downstream ERP.

Do: Add a “rate limit” node at the start keyed by user_id; over N/min → queue.

Pattern 7 — Money values must be sourced#

Problem: user asks refund amount; LLM cannot compute it.

Do:

1. HTTP(order API) → order_amount
2. LLM → answer using `{{order_amount}}`
3. Prompt: "Use `{{order_amount}}` exactly; do not rewrite."

Pattern 8 — Human fallback#

Problem: AI can’t answer — don’t dead-end.

Do: failure branch calls Chatwoot API to open a ticket with the full context as a note.

Pattern 9 — Multi-turn form#

Problem: collecting name / email / description rarely happens in one turn.

Do: Use Conversation App, check completeness per turn, ask for whatever is missing.

Pattern 10 — A/B testing prompts#

Problem: was the new prompt actually better?

Do: random branch — 50% old prompt, 50% new — tag user_id, compare CSAT later.

General lessons#

• Keep workflow nodes atomic — easier to debug individually
• Declare key variables in the Start node, not scattered
• Use different temperatures for different LLM nodes (generation vs classification)
• Use Dify’s debug pane to step through — faster than full logs