10 common pitfalls when self-hosting open-source AI support (with fixes)

The 10 most frequent issues we've seen rolling out Chatwoot + Dify / FastGPT over the past year — repro and fix for each.

1. Chatwoot can’t send mail#

Symptom: signup / ticket notification emails fail.

Cause: default dev SMTP; MAILER_SENDER_EMAIL mismatched with the verified domain.

Fix: Postmark or SES, ensure MAILER_SENDER_EMAIL is in the sender domain’s SPF / DKIM.

Symptom: chat replies “I don’t have relevant info” after you uploaded docs.

Cause: 90% of the time it’s an embedding-dimension mismatch — you switched models without resetting the vector store.

Fix: rebuild the KB, or pin the embedding model in Settings.

Symptom: replies reference your internal SOP.

Cause: internal and external docs share one dataset.

Fix: split into two datasets. Prompt: “Only cite content from dataset ‘public’.“

Cause: container listens on 127.0.0.1 rather than 0.0.0.0.

Fix: update docker-compose ports or bind to 0.0.0.0.

Cause: chunking strategy is re-picked each time; new docs dilute high-quality old ones.

Fix: pin chunking, build new docs in a side dataset, evaluate, then merge.

Symptom: AI forgets prior turns after 5 messages.

Cause: Dify default context window is 5 turns; models have hard limits too.

Fix: bump Dify to 15-20 turns or add a summarization node.

Symptom: 8-15 seconds first response.

Cause: streaming not enabled + Cloudflare buffering + wrong model.

Fix: enable streaming, turn off Auto Minify / Rocket Loader, switch to low-latency model (Doubao / Groq Llama).

Cause: tool schema validation fails silently.

Fix: hit the tool endpoint with Postman first, verify the OpenAPI schema matches Dify exactly.

Cause: prompt missing constraints + no pre-filter.

Fix: list forbidden topics in prompt + add a keyword pre-filter (n8n node).

Symptom: LLM tokens 3-5× normal.

Cause: workflow loop or bot traffic spam.

Fix: