flag92 flag92
Benchmarks · practice · thinking

Blog

Notes on choosing, deploying and operating open-source AI support.

Wed Mar 25 2026 08:00:00 GMT+0800 (中国标准时间) ·
selectionChatwootDify

Chatwoot Captain or bolt-on Dify? Choosing your AI brain

Chatwoot's built-in Captain can reply with an LLM directly — but it trades off against a bolted-on Dify Agent Bot. Clear criteria here.

Wed Mar 18 2026 08:00:00 GMT+0800 (中国标准时间) ·
pitfallspractice

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.

Wed Feb 25 2026 08:00:00 GMT+0800 (中国标准时间) ·
deep-diveFastGPTRAG

Tuning FastGPT KBs — from 0.62 to 0.89 accuracy, step by step

FastGPT defaults give ~60% retrieval accuracy. Tuned right it reaches ~90%. Step-by-step measurements for chunking, augmentation, reranking.

Wed Feb 18 2026 08:00:00 GMT+0800 (中国标准时间) ·
deep-divelocal inferenceOpen WebUI

Fully local AI support — Open WebUI + Ollama + Qwen on one GPU box

Under strict no-data-egress mandates, can one GPU server run the entire AI support stack? Hardware, models, KB, exposure — end to end.

Wed Feb 11 2026 08:00:00 GMT+0800 (中国标准时间) ·
deep-diven8nautomation

n8n as central orchestration for AI support — 10 real automations

The business-logic layer of an AI support stack gets ignored. n8n's 800+ nodes glue Chatwoot, Dify, CRM, ERP and Slack together.

Wed Feb 04 2026 08:00:00 GMT+0800 (中国标准时间) ·
deep-diveRAGFlowDify

RAGFlow vs Dify KBs — head-to-head on one complex PDF

A 30-page product manual with tables, two-column layout and scanned pages. Indexed with both. Top-5 accuracy and answer quality compared.

Wed Apr 29 2026 08:00:00 GMT+0800 (中国标准时间) ·
opsKPImetrics

AI support KPI framework — how to prove the AI actually saved money

Don't measure deflection alone — here's a 4-layer 12-metric framework covering experience, cost and training.

Wed Apr 22 2026 08:00:00 GMT+0800 (中国标准时间) ·
comparisonDifyFastGPT

Dify vs FastGPT — 8 dimensions to decide

The two most popular open-source LLM app platforms in China overlap heavily. This compares them across 8 dimensions and gives recommendations.

Wed Apr 15 2026 08:00:00 GMT+0800 (中国标准时间) ·
practiceRAGknowledge base

KB preparation in practice — from Notion / Lark / Confluence to RAG

A bad KB ruins any LLM budget. End-to-end repeatable pipeline — export, clean, chunk, augment, evaluate — with Python.

Wed Apr 08 2026 08:00:00 GMT+0800 (中国标准时间) ·
benchmarkRAGquality

How to evaluate your RAG quality — 5 metrics, 3 toolkits

"Feels better than last week" is not measurement. Five quantitative RAG metrics and three open-source evaluators.

Wed Apr 01 2026 08:00:00 GMT+0800 (中国标准时间) ·
deploymentCloudflareops

Production-grade Cloudflare Tunnel tuning for AI support

One-command Cloudflare Tunnel for Chatwoot + Dify hides production gotchas — WebSocket drops, timeouts, monitoring blind spots. A checklist.

Tue May 12 2026 08:00:00 GMT+0800 (中国标准时间) ·
benchmarkembeddingRAG

2026 Chinese embedding benchmark — bge-m3, Conan, m3e, bce, OpenAI

Same Chinese KB, same real support questions — five embedding models compared on retrieval accuracy, speed and cost.

Tue Mar 10 2026 08:00:00 GMT+0800 (中国标准时间) ·
deep-diveChatwoot

Chatwoot deep review — 6 months in production, the unvarnished take

We ran Chatwoot 4.x in production for 6 months — 40k+ tickets, 12 agents, full debrief on Inbox, Captain, API and gotchas.

Tue Mar 03 2026 08:00:00 GMT+0800 (中国标准时间) ·
deep-diveDifyworkflow

Dify workflow patterns — 10 we actually use in production

Ten Dify workflow patterns that survive production — KB fallback, context compression, concurrent tool calls, retries and more.

Sun May 10 2026 08:00:00 GMT+0800 (中国标准时间) ·
thinkingevaluation

Why self-hosting an AI support stack pays off in 2026

SaaS support keeps getting pricier while open-source matures. A framework across cost, compliance and control.

Sun Jan 25 2026 08:00:00 GMT+0800 (中国标准时间) ·
migrationChatwootZendesk

Migrating from Zendesk to Chatwoot — translating tickets and triggers

Zendesk's ticketing is heavier than Intercom's; the hardest part of migrating to Chatwoot is rewriting Triggers / Automations / Macros.

Sun Jan 18 2026 08:00:00 GMT+0800 (中国标准时间) ·
migrationDifyOpenAI

Migrating from OpenAI GPTs to Dify — bring your Custom GPT home

Custom GPTs are great for tinkering but limit you when you go commercial, need an API or want to swap models. Step-by-step Builder → Dify migration.

Sun Feb 01 2026 08:00:00 GMT+0800 (中国标准时间) ·
migrationChatwootIntercom

Migrating from Intercom to Chatwoot — the complete path

Data export, field mapping, parallel running, cutover — a battle-tested playbook with scripts and traps.

Fri May 15 2026 08:00:00 GMT+0800 (中国标准时间) ·
case studye-commercepractice

Case · A cross-border e-commerce running 80k daily orders — 6 months on Chatwoot + Dify

A Shopify + WhatsApp e-commerce shop migrated from Intercom to Chatwoot + Dify. Six months of real numbers — cost, deflection, satisfaction.

Fri May 08 2026 08:00:00 GMT+0800 (中国标准时间) ·
case studymigrationSaaS

Case · A B2B SaaS migrated from Intercom to Chatwoot in 6 weeks

A 2,000-customer, $2M ARR B2B SaaS completed Intercom → Chatwoot migration in 6 weeks. Full timeline, headcount, scars.

Fri May 08 2026 08:00:00 GMT+0800 (中国标准时间) ·
benchmarkLLMmodel selection

2026 Chinese-support LLM bake-off — Qwen, DeepSeek, GLM, Doubao, ERNIE

Same support prompt and knowledge base — which of the five China-trained LLMs ships the best AI support? 200 real questions decide.

Fri May 01 2026 08:00:00 GMT+0800 (中国标准时间) ·
case studyfinancecompliance

Case · A securities firm shipped a fully local AI support — the compliance path

No data egress, regulator approval, internal + external audit, DR plan. The real compliance journey for a mid-sized securities brokerage's local AI support.

Fri Apr 24 2026 08:00:00 GMT+0800 (中国标准时间) ·
case studyeducationpromo day

Case · An online-edu promo day with 50× traffic — engineering notes

A coding-education company handled 120k conversations in 24 hours during 618 sale, AI deflection 81%. How engineering pre-tested, rate-limited, degraded gracefully.

Fri Apr 17 2026 08:00:00 GMT+0800 (中国标准时间) ·
case studyhealthcarecompliance

Case · How a healthcare provider built RAG with zero PHI leakage

An internet hospital launched AI health consultations. The hardest part — PHI must never reach the LLM. Their 3-tier KB + Pipeline redaction + audit architecture.

Fri Apr 10 2026 08:00:00 GMT+0800 (中国标准时间) ·
case studyopen sourceDiscord

Case · An open-source project handled a 300k Discord community with Chatwoot + Dify

A 18k-star OSS project went from drowning in Discord to clean AI + volunteer mod tiers handling 300k community asks.

Search

Press ⌘ K to open