WFGY/ProblemMap/context-drift.md
2025-07-28 10:27:02 +08:00

100 lines
2.9 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# 🧠 Problem: Long QA Chains Drift Off-Topic
### 📍Context
Even when each individual response is locally correct, many AI agents begin to **semantically drift** as question-answer chains grow longer.
Symptoms include:
- Subtle shifts in topic over 510 turns
- Forgotten user goals
- Misalignment between early and late context
- The agent redefines the question mid-conversation
---
## 🚨 Why Traditional RAG Fails Here
| Weakness | Description |
|----------|-------------|
| No persistent memory | Most systems treat each QA turn as an isolated prompt context |
| Embedding overlap is fragile | Token overlap does not equal topic stability |
| No tracking of concept flow | Systems cant trace how topics evolved or when they “jumped” |
---
## ✅ WFGY Solution
WFGY uses **semantic delta tracking** and **Tree-based memory nodes** to detect and prevent drift.
### 1. Semantic Tree Memory
- Each major concept shift is recorded as a node
- You can view and backtrack logic flow across topics
### 2. ΔS as Drift Detector
- When new input diverges from past nodes (ΔS > 0.6), the system logs a new branch
- This allows structured topic separation and detection of "semantic fatigue"
### 3. λ_observe Vector
- Flags if the reasoning is now divergent or chaotic
- Helps model decide whether to re-anchor or warn the user
---
## 🛠 How to Use in TXT OS
```txt
Step 1 — Start the console
> Start
Step 2 — Ask a sequence of loosely connected questions:
> "What is the policy on returns?"
> "And if it's a gift item?"
> "Now, what about shipping zones?"
> "What if I'm in another country?"
Step 3 — Type `view` to inspect the Tree
Youll see:
- Nodes logged with ΔS and λ_observe
- Clear detection of topic shifts
- Logic branching when context drift occurs
````
---
## 🔬 Example Output
```txt
* Topic: Gift Return Policy | ΔS: 0.22 | λ: → | Module: BBMC
* Topic: International Shipping | ΔS: 0.74 | λ: ← | Module: BBPF, BBCR
```
The system realized a **new conceptual frame** was entered and recorded the shift accordingly.
---
## 🔗 Related Modules
* `BBMC` — Identifies when the concept anchor has shifted
* `BBPF` — Supports divergent paths while maintaining logic
* `BBCR` — May reroute reasoning or pause to prevent collapse
* `Semantic Tree` — Memory structure to prevent context loss
---
## 📌 Status
| Feature | Status |
| -------------------- | ----------------------------------- |
| Tree node logging | ✅ stable |
| ΔS-based topic split | ✅ working |
| λ\_observe awareness | ✅ working |
| Auto recall or warn | ⚠️ partial (manual inspect for now) |
---
## ✍️ Summary
WFGY doesn't just answer — it remembers why you're asking.
If you're tired of long chats forgetting your intent, this is the solution layer you're missing.