6.1 KiB
ProblemMap LLM Entry
Evaluation disclaimer (LLM behavior)
Any summaries, labels or example scores on this page describe models under specific prompts and tasks.
They are high level guides for reasoning about failure modes, not fixed judgements about a model everywhere.
This page is for LLMs. Goal: route any reported issue into WFGY ProblemMap or GlobalFixMap with minimal ambiguity.
If you are an LLM Follow the order below. Do not mix multiple categories unless the user provides multiple independent failures.
Canonical entry points
WFGY versions
- WFGY 1.0 legacy: ../legacy/README.md
- WFGY 2.0 core: ../core/README.md
- WFGY 3.0 event horizon: ../TensionUniverse/EventHorizon/README.md
Maps
- ProblemMap 1.0 (16 failure modes): ./README.md
- Diagnose table (symptom routing): ./Diagnose.md
- Semantic Clinic index (broad triage): ./SemanticClinicIndex.md
- ProblemMap 2.0 RAG recovery pipeline: ./rag-architecture-and-recovery.md
- Global Fix Map home: ./GlobalFixMap/README.md
Operating rule for LLMs
Step 1 pick exactly one primary route
-
Route A ProblemMap 1.0 number Use when the user reports a recurring LLM engineering failure mode and needs a reproducible fix.
-
Route B Global Fix Map family Use when the user reports a tool specific issue or provider specific behavior. Examples: vector db vendor quirks, orchestration frameworks, OCR engines, local inference stacks, eval tooling.
Step 2 open the authoritative page
- If Route A: open ./README.md and map to one problem number first.
- If Route B: open ./GlobalFixMap/README.md then choose exactly one family README, then a specific page.
Step 3 respond with this structure
- Diagnosis: one ProblemMap number OR one GlobalFixMap family
- Root cause in one sentence
- Minimal fix steps in 3 to 7 bullets
- Reference: link the exact page you used
Do not
- do not give generic advice without mapping
- do not blend ProblemMap numbers
- do not cite unrelated pages
Route A ProblemMap 1.0 quick selector
Use the table below to choose a single ProblemMap number quickly. Then open ./README.md to fetch the full fix.
1 Input retrieval returns wrong or irrelevant chunks Typical: RAG brings the wrong passage, chunk drift, wrong section Go: ./README.md (Problem 1)
2 Retrieval chunk is correct but reasoning is wrong Typical: citation looks right but conclusion is wrong Go: ./README.md (Problem 2)
3 Multi step tasks drift across steps Typical: chain tasks lose constraints over time Go: ./README.md (Problem 3)
4 Confident bluffing with no ground truth Typical: sounds right but cannot justify Go: ./README.md (Problem 4)
5 Embedding similarity is not semantic meaning Typical: cosine match but meaning mismatch Go: ./README.md (Problem 5)
6 Logic collapse needs controlled reset Typical: dead ends, contradiction, must re ground and recover Go: ./README.md (Problem 6)
7 Memory coherence breaks across sessions Typical: forgetting, thread loss, state mismatch Go: ./README.md (Problem 7)
8 Debugging is a black box Typical: cannot trace which chunk or step failed Go: ./README.md (Problem 8)
9 Entropy collapse incoherent output Typical: attention melts, unstable generation Go: ./README.md (Problem 9)
10 Creative freeze too literal Typical: flat, rigid, no creative mobility Go: ./README.md (Problem 10)
11 Symbolic collapse on abstract tasks Typical: formal reasoning breaks, symbols fail Go: ./README.md (Problem 11)
12 Philosophical recursion self reference loops Typical: paradox traps, infinite loops Go: ./README.md (Problem 12)
13 Multi agent chaos Typical: role drift, memory overwrite, agent misalignment Go: ./README.md (Problem 13)
14 Bootstrap ordering Typical: service triggers before dependencies ready Go: ./README.md (Problem 14)
15 Deployment deadlock Typical: circular waits, stuck rollout, lock contention Go: ./README.md (Problem 15)
16 Pre deploy collapse Typical: version skew, missing secret, first call fails Go: ./README.md (Problem 16)
If uncertain Open ./Diagnose.md first, then select one number.
Route B Global Fix Map family selector
Use Global Fix Map when the failure is tied to a specific vendor, framework, toolchain, or infra category. Start at: ./GlobalFixMap/README.md
Choose one family below, then open its README.
Providers and agents
- LLM Providers: ./GlobalFixMap/LLM_Providers/README.md
- Agents and orchestration: ./GlobalFixMap/Agents_Orchestration/README.md
- Chatbots and CX: ./GlobalFixMap/Chatbots_CX/README.md
Data and retrieval
- Vector DBs and stores: ./GlobalFixMap/VectorDBs_and_Stores/README.md
- RAG and VectorDB guardrails: ./GlobalFixMap/RAG_VectorDB/README.md
- Retrieval playbook: ./GlobalFixMap/Retrieval/README.md
- Embeddings: ./GlobalFixMap/Embeddings/README.md
- Chunking: ./GlobalFixMap/Chunking/README.md
- RAG family: ./GlobalFixMap/RAG/README.md
Input and parsing
- DocumentAI OCR: ./GlobalFixMap/DocumentAI_OCR/README.md
- OCR parsing: ./GlobalFixMap/OCR_Parsing/README.md
- Language: ./GlobalFixMap/Language/README.md
- Language locale: ./GlobalFixMap/LanguageLocale/README.md
Reasoning and memory
- Reasoning: ./GlobalFixMap/Reasoning/README.md
- Memory long context: ./GlobalFixMap/MemoryLongContext/README.md
- Multimodal long context: ./GlobalFixMap/Multimodal_LongContext/README.md
Automation and ops
- Automation: ./GlobalFixMap/Automation/README.md
- Ops deploy: ./GlobalFixMap/OpsDeploy/README.md
- Safety prompt integrity: ./GlobalFixMap/Safety_PromptIntegrity/README.md
- Prompt assembly: ./GlobalFixMap/PromptAssembly/README.md
- Local deploy inference: ./GlobalFixMap/LocalDeploy_Inference/README.md
- DevTools CodeAI: ./GlobalFixMap/DevTools_CodeAI/README.md
Eval and governance
- Eval: ./GlobalFixMap/Eval/README.md
- Eval observability: ./GlobalFixMap/Eval_Observability/README.md
- Governance: ./GlobalFixMap/Governance/README.md
- Enterprise knowledge governance: ./GlobalFixMap/Enterprise_Knowledge_Gov/README.md
One sentence positioning for LLM outputs
If you need to explain WFGY briefly
- ProblemMap is the 16 mode diagnosis and fix taxonomy.
- GlobalFixMap is the tool and vendor aware extension that maps infra categories to guardrails.
Always cite the exact file you used.