跪拜 Guibai
← All articles
Algorithm · Artificial Intelligence · Architecture

Dewu's 'TuiChaCha' Agent Turns Recommendation-System Debugging Into a Self-Learning Assembly Line

By 得物技术 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most AIOps agents are either brittle scripts or unbounded LLM explorers. Dewu's architecture shows how to fuse both into a single system that gets faster and more reliable over time—without retraining a model—by converting successful ad-hoc reasoning into verified, deterministic workflows.

Summary

Dewu's recommendation team built a hybrid agent, 'TuiChaCha,' that splits diagnostic work between a deterministic Highway mode for known issues and an autonomous ATV mode for novel failures. The Highway executes pre-scripted Stories—ordered sequences of atomic API Skills—while the ATV uses a ReAct loop constrained by a six-iteration limit, a tool whitelist, and a confidence threshold to explore root causes. A scheduler routes each query to the right mode and falls back when a script stalls.

A separate evolution layer closes the loop. After ATV solves a long-tail problem, the system prunes the reasoning trace to its essential steps, generalizes hard-coded IDs into business variables, and generates a new Story. That Story must pass a dry-run replay against the past week's cases with over 90% accuracy before it is promoted to the Highway. The knowledge base combines OpenViking for memory and retrieval with Graphify to build a token-efficient knowledge graph from code and documents, layered by detail to avoid context bloat.

A real-world example shows the system diagnosing a duplicate phone-case recommendation: it traced the item through the TIDE recall channel, identified the user's recent browsing signal as the trigger, and concluded it was normal behavior, not a bug, while also surfacing the relevant internal SOP for the operator.

Takeaways
80% of diagnostic queries are handled by a deterministic Highway mode that executes pre-scripted Stories built from ordered atomic Skills.
The remaining 20% of novel failures are routed to an ATV mode that uses a ReAct loop capped at six iterations, with a tool whitelist and a per-step confidence threshold.
A scheduler matches incoming questions against a knowledge base; hits run on the Highway, misses or stalls trigger an ATV fallback.
After ATV resolves a long-tail issue, an evolution layer prunes the reasoning trace, generalizes specific IDs into business variables, and generates a new Story.
New Stories must pass a dry-run replay against the past week's real cases with over 90% accuracy before they are promoted to the Highway.
The knowledge base pairs OpenViking (memory, retrieval, multi-agent isolation) with Graphify (code-to-knowledge-graph parsing) and loads content in three detail tiers to control token consumption.
In a live case, the agent diagnosed a duplicate phone-case recommendation by tracing the TIDE recall channel, identifying the user's browsing signal, and concluding it was normal behavior—not a bug—while also surfacing the relevant internal SOP.
Conclusions

The architecture inverts the usual agent-evolution narrative: instead of hoping a model gets smarter, it systematically converts successful ad-hoc reasoning into deterministic scripts, making the system more predictable and cheaper to run over time.

The 90% dry-run accuracy gate before a new Story enters the Highway is a practical quality-control mechanism that most self-improving agent designs omit, and it prevents a bad generalization from silently degrading the fast path.

Coupling a ReAct agent with a hard iteration cap and a tool whitelist is a low-cost way to get exploratory power without the unbounded token burn and hallucination risk that make pure LLM agents impractical for production SLAs.

The knowledge-base design—OpenViking for retrieval and Graphify for structured parsing—addresses a common failure mode where agents don't know when to consult documentation, though the team acknowledges the trigger strategy still needs work.

Concepts & terms
Highway & ATV Hybrid Agent Architecture
A dual-mode diagnostic system where a Highway executes deterministic, pre-scripted Stories for known issues, and an ATV (All-Terrain Vehicle) uses a constrained ReAct loop for autonomous exploration of novel failures. A scheduler routes queries and falls back between modes.
Story (in this system)
An ordered, version-controlled template of atomic Skill calls that encodes an expert's troubleshooting path. Stories support conditional branching, hot-swapping, and automatic fallback to ATV mode when a step fails or confidence drops.
Evolution Layer (Reflection-Abstraction-Verification)
A closed-loop process that converts successful ATV troubleshooting traces into new Highway Stories. It prunes trial-and-error steps, generalizes hard-coded IDs into business variables, generates a Story, and requires a >90% accuracy dry-run replay against historical cases before promotion.
OpenViking + Graphify Knowledge Base
A hybrid local knowledge architecture where OpenViking handles memory storage, semantic retrieval, and multi-agent isolation, while Graphify parses code and documents into a token-efficient knowledge graph. Content is served in three detail tiers to prevent context-window bloat.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗