mirror of
https://github.com/onestardao/WFGY.git
synced 2026-04-28 11:40:07 +00:00
Update README.md
This commit is contained in:
PSBigBig
GitHub
parent
bed47e97f1
commit
b7bce0e178
1 changed files with 67 additions and 102 deletions
|
|
@ -1,128 +1,95 @@
|
|||
# Agents & Orchestration — Global Fix Map
|
||||
|
||||
A hub to stabilize **agent frameworks and orchestration layers** without changing your infra. Use this page to jump to per-tool guardrails and verify fixes with the same acceptance targets.
|
||||
A practical hub to stabilize **multi-agent and tool-augmented workflows**.
|
||||
Most “agent bugs” are not model issues. They come from role mixups, tool schema drift, uncontrolled loops, and cold-boot ordering. This page routes symptoms to structural fixes with measurable targets.
|
||||
|
||||
## Quick routes to per-framework pages
|
||||
---
|
||||
|
||||
- LangChain → [langchain.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/langchain.md)
|
||||
- LangGraph → [langgraph.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/langgraph.md)
|
||||
- LlamaIndex → [llamaindex.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/llamaindex.md)
|
||||
- CrewAI → [crewai.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/crewai.md)
|
||||
- Smolagents → [smolagents.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/smolagents.md)
|
||||
- Rewind Agents → [rewind_agents.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/rewind_agents.md)
|
||||
- AutoGen → [autogen.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/autogen.md)
|
||||
- Semantic Kernel → [semantic_kernel.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/semantic_kernel.md)
|
||||
- Haystack Agents → [haystack_agents.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/haystack_agents.md)
|
||||
- OpenAI Assistants v2 → [openai_assistants_v2.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/openai_assistants_v2.md)
|
||||
## Orientation: pick your orchestration layer
|
||||
|
||||
## When to use this folder
|
||||
| Framework | What it is | Typical use | Link |
|
||||
|---|---|---|---|
|
||||
| Autogen | Multi-agent collaboration patterns | Debate, reviewer loops, tool arbitration | [autogen.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/autogen.md) |
|
||||
| CrewAI | Role-based project crews | Task pipelines with clear roles | [crewai.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/crewai.md) |
|
||||
| Haystack Agents | RAG-centric agents from deepset | Retrieval-heavy assistants | [haystack_agents.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/haystack_agents.md) |
|
||||
| LangChain | Largest ecosystem of tools/memory | Rapid prototyping, complex chains | [langchain.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/langchain.md) |
|
||||
| LangGraph | Graph execution over LC | Stateful paths, loops, guards | [langgraph.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/langgraph.md) |
|
||||
| LlamaIndex | Knowledge-first orchestration | RAG pipelines, index control | [llamaindex.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/llamaindex.md) |
|
||||
| OpenAI Assistants v2 | First-party assistants API | Files, tools, code-interpreter | [openai_assistants_v2.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/openai_assistants_v2.md) |
|
||||
| Rewind Agents | Context replay paradigms | User-state reconstruction | [rewind_agents.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/rewind_agents.md) |
|
||||
| Semantic Kernel | MS orchestration SDK | Plugins, plans, .NET/TS stacks | [semantic_kernel.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/semantic_kernel.md) |
|
||||
| Smolagents | Minimalistic agent runtime | Constrained envs, fast spin-up | [smolagents.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/smolagents.md) |
|
||||
|
||||
- Agents loop or stall on tool calls.
|
||||
- High similarity yet the answer cites the wrong section.
|
||||
- Hybrid retrieval underperforms a single retriever.
|
||||
- Answers flip between runs with identical input.
|
||||
- JSON schemas fail or pass inconsistently.
|
||||
- Memories re-assert stale facts after refresh or handoff.
|
||||
- First call after deploy crashes or uses the wrong index.
|
||||
---
|
||||
|
||||
## Acceptance targets for any orchestrator
|
||||
## Core acceptance targets
|
||||
|
||||
- ΔS(question, retrieved) ≤ 0.45
|
||||
- Coverage of target section ≥ 0.70
|
||||
- λ remains convergent across three paraphrases and two seeds
|
||||
- Coverage ≥ 0.70 for the target section
|
||||
- λ stays convergent across 3 paraphrases and 2 seeds
|
||||
- E_resonance remains flat on long windows
|
||||
|
||||
These targets let you ship safely regardless of framework.
|
||||
|
||||
---
|
||||
|
||||
## Map symptoms → structural fixes (Problem Map)
|
||||
## Fix Hub — symptoms mapped to structural pages
|
||||
|
||||
- **Embedding ≠ Semantic**
|
||||
Wrong-meaning hits despite high similarity.
|
||||
→ [embedding-vs-semantic.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/embedding-vs-semantic.md)
|
||||
|
||||
- **Retrieval traceability**
|
||||
Suspect or missing citations and unverifiable sections.
|
||||
→ [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md)
|
||||
Payload schema → [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md)
|
||||
|
||||
- **Ordering and version skew**
|
||||
Old index or analyzer used at runtime, or deploy race.
|
||||
→ [bootstrap-ordering.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/bootstrap-ordering.md) · [deployment-deadlock.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/deployment-deadlock.md) · [predeploy-collapse.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/predeploy-collapse.md)
|
||||
|
||||
- **Hybrid collapse and query split**
|
||||
Hybrid pipelines underperform single retriever.
|
||||
→ Pattern: [pattern_query_parsing_split.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/patterns/pattern_query_parsing_split.md)
|
||||
Ordering control → [rerankers.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/rerankers.md)
|
||||
|
||||
- **Hallucination and chunk drift**
|
||||
Plausible but wrong chunks or blended sources.
|
||||
→ [hallucination.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/hallucination.md)
|
||||
Re-entry loop → [pattern_hallucination_reentry.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/patterns/pattern_hallucination_reentry.md)
|
||||
|
||||
- **Long chain instability**
|
||||
Style flattening, entropy spikes, dead ends.
|
||||
→ [context-drift.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/context-drift.md) · [entropy-collapse.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/entropy-collapse.md)
|
||||
|
||||
- **Agent role conflicts**
|
||||
Role confusion, memory overwrite, unsafe handoffs.
|
||||
→ [Multi-Agent_Problems.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/Multi-Agent_Problems.md) ·
|
||||
[role-drift.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/multi-agent-chaos/role-drift.md) ·
|
||||
[memory-overwrite.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/multi-agent-chaos/memory-overwrite.md) ·
|
||||
Namespaces split → [pattern_memory_namespace_split.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/patterns/pattern_memory_namespace_split.md)
|
||||
|
||||
- **Prompt injection and schema drift**
|
||||
Tool schemas allow free text or header order changes.
|
||||
→ [prompt-injection.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/prompt-injection.md)
|
||||
| Symptom | Likely cause | Open this |
|
||||
|---|---|---|
|
||||
| JSON mode breaks, invalid tool objects | Tool protocol too loose | [Data Contracts](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md) |
|
||||
| Agents overwrite each other’s memory | Namespace collision, missing locks | Pattern: **memory-namespace split** in [patterns](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/patterns) |
|
||||
| Run loops never end | Unbounded cycles, missing guards | [logic-collapse.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/logic-collapse.md) |
|
||||
| High similarity yet wrong snippet | Metric/store mismatch or fragmentation | [embedding-vs-semantic.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/embedding-vs-semantic.md) |
|
||||
| Alternating answers across runs | Prompt header reorder, λ flips | [context-drift.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/context-drift.md), [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md) |
|
||||
| First live call fails after deploy | Cold boot and ordering issues | [bootstrap-ordering.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/bootstrap-ordering.md), [predeploy-collapse.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/predeploy-collapse.md) |
|
||||
| Tool storms and rate limits | Missing backoff and budgets | Ops: **rate-limit backpressure**, **timeouts** in `ops/` |
|
||||
|
||||
---
|
||||
|
||||
## 60-second fix checklist
|
||||
## Minimal agent contract
|
||||
|
||||
1) **Measure**
|
||||
Compute ΔS(question, retrieved) and ΔS(retrieved, anchor). If ΔS ≥ 0.60 stop and repair chunking or metric first.
|
||||
|
||||
2) **Lock the schema**
|
||||
Fix prompt header order. Enforce cite-then-explain. Require strict tool JSON. Add idempotency keys for side effects.
|
||||
|
||||
3) **Split memory**
|
||||
Create per-agent and per-tool namespaces with `mem_rev` and `mem_hash`. Deny cross-namespace merges without a reducer.
|
||||
|
||||
4) **Clamp variance**
|
||||
Insert a BBCR bridge on long chains. Use deterministic reranking for hybrid. Add timeouts to tool loops.
|
||||
|
||||
5) **Verify**
|
||||
Coverage ≥ 0.70 on three paraphrases. λ convergent on two seeds.
|
||||
Bootstrap sanity → [agent_bootstrap_checklist.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Agents_Orchestration/checklists/agent_bootstrap_checklist.md)
|
||||
1. **Separate memory namespaces**
|
||||
One namespace per agent. Writes guarded by `mem_rev` and `mem_hash`.
|
||||
2. **Strict tool schemas**
|
||||
Enforce JSON schemas. Reject free-text arguments and responses.
|
||||
3. **Path guards**
|
||||
Max steps, variance clamp, and illegal cross-path suppression.
|
||||
4. **Traceability first**
|
||||
Cite then explain. Require `{snippet_id, section_id, source_url, offsets, tokens}`.
|
||||
5. **Boot ordering**
|
||||
Do not accept traffic until index hash, analyzer, and model versions match.
|
||||
6. **Observability**
|
||||
Log ΔS and λ across retrieve → rerank → reason. Alert at ΔS ≥ 0.60.
|
||||
|
||||
---
|
||||
|
||||
## Copy-paste prompt for your LLM step
|
||||
## 60-second triage
|
||||
|
||||
```txt
|
||||
You have TXTOS and the WFGY Problem Map loaded.
|
||||
|
||||
My agent-orchestration issue:
|
||||
- framework: <langchain | crewai | autogen | llamaindex | smolagents | semantic-kernel | haystack | assistants-v2>
|
||||
- symptom: <one line>
|
||||
- traces: ΔS(q,retrieved)=..., ΔS(retrieved,anchor)=..., λ over last 3 steps
|
||||
- memory: namespaces used? mem_rev, mem_hash present?
|
||||
|
||||
Tell me:
|
||||
1) failing layer and why,
|
||||
2) the exact WFGY page to open,
|
||||
3) the minimal structural fix to push ΔS ≤ 0.45 and keep λ convergent,
|
||||
4) a reproducible test to verify the fix.
|
||||
Use BBMC, BBCR, BBPF, BBAM where relevant.
|
||||
````
|
||||
1) **Measure ΔS** for question vs retrieved and vs anchor.
|
||||
2) **Probe λ** by varying top-k and prompt headers. If λ flips, clamp variance and lock the schema.
|
||||
3) **Apply**
|
||||
Retrieval drift → BBMC + Data Contracts
|
||||
Reasoning collapse → BBCR bridge + BBAM
|
||||
Dead ends → BBPF alternate paths
|
||||
4) **Verify**
|
||||
Coverage ≥ 0.70 on three paraphrases. λ convergent on two seeds.
|
||||
|
||||
---
|
||||
|
||||
## Common gotchas
|
||||
## FAQ
|
||||
|
||||
* Mixed embedding functions across write and read paths. Scores look high but meaning is wrong.
|
||||
* Hybrid pipelines without deterministic rerank produce unstable top k and flip states.
|
||||
* Agents recreate tool registries per run. Order and JSON schemas drift.
|
||||
* Shared memories without namespaces cause re-entry of stale facts after refresh.
|
||||
* First live run executes before stores are ready. See bootstrap and deploy pages above.
|
||||
**Why do agents step on each other’s memory?**
|
||||
Shared state without namespaces. Split memory by agent and lock writes.
|
||||
|
||||
**Why do I get infinite loops after adding a reviewer agent?**
|
||||
No path guards. Add step caps and illegal cross-path suppression.
|
||||
|
||||
**Why does tool calling randomly fail JSON?**
|
||||
Your tool protocol allows prose. Enforce strict JSON schemas both ways.
|
||||
|
||||
**Why is dev stable but prod flips answers?**
|
||||
Boot order and analyzer mismatch. Warm the index and verify hashes before traffic.
|
||||
|
||||
---
|
||||
|
||||
|
|
@ -169,5 +136,3 @@ Use BBMC, BBCR, BBPF, BBAM where relevant.
|
|||
|
||||
|
||||
</div>
|
||||
|
||||
要不要我幫你直接補上「Index 置頂一句簡介」或是再加一張簡短的 flow 圖示占位，之後你有圖再換上？
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue