The Trace is an observability platform for LLMs, self-hosted, specifically for bots that are running in production in Vietnam. The product helps operations teams clearly see each interaction of the bot: incoming messages, NLU processing, answer generation, and responses sent to users.

The problem that Trace solves: Vietnamese bots on Zalo, Lark, or Teams are often a "black box" — it's hard to know why the bot gives incorrect answers, how much model cost is incurred, or where the error lies. Trace brings the entire processing flow to light for easier debugging and optimization.

Main features:

• Multi-channel connections: Zalo OA, Zalo Bot, FPT.AI Conversation, Viettel ASR/TTS, Lark, Google Chat, and Teams.
• Channel (webhooks) for receiving incoming messages or integration via SDK/OTLP.
• A mapping pathway for each conversation: inbound → NLU → generation → reply.
• A console built on Langfuse OSS v4.33 with a Vietnamese interface: traces, sessions, scores, prompts, and model costs.

For whom: technical and operations teams building and running Vietnamese chatbot/voicebot at production scale and need to monitor quality as well as costs.

How to get started: run locally using Docker — the console at port :3000, channels and /hooks at port :43173. A demo project called Bot Zalo shop is available for immediate experience.

Difference: self-hosted (conversation data resides within your infrastructure) and optimized for the Vietnamese ecosystem — Zalo, FPT.AI, Viettel — with a Vietnamese interface, something international observability tools do not directly support.