Add cross-project pattern registry for retrieval-augmented generalization

- Create patterns/ directory with README, manifest, and 10 initial patterns
  covering Ollama JSON fallback, API escaping, deprecation, PTY auth,
  queue-poll, LLM-as-parser, credential rotation, reverse proxy binding,
  human approval gates, and transient retry.
- Wire pattern loading into architecture/pipeline.js based on task tags.
- Update architecture/orchestrator.js to load patterns and surface them in
  the system prompt.
- Update MEMORY.md, ARCHITECTURE.md, and CONTEXT.md to document the registry
  and record the decision.
This commit is contained in:
JC Beasley
2026-08-06 12:46:09 -07:00
parent d8989f9371
commit 0bd719ab91
17 changed files with 676 additions and 32 deletions
+33
View File
@@ -0,0 +1,33 @@
# Pattern: LLM-as-Parser with Structured Fallback
## Symptom
A workflow asks an LLM to parse, classify, or extract information from unstructured input. Sometimes the output is correct but poorly formatted; sometimes it is wrong or inconsistent. Downstream nodes cannot rely on it without a validation step.
## Affected Projects
- Cyber Tips Newsletter pipeline
- Proxmox VE snapshot summarizer
- Email inbox triage agent
- Monthly IT newsletter generator
## Root Cause
Using an LLM as a parser combines the power of fuzzy reasoning with the fragility of probabilistic output. Without a structured fallback, the pipeline is brittle.
## Standard Fix
1. Ask the LLM for structured output (JSON/schema) when possible.
2. Add a validation layer that checks required fields and value ranges.
3. Add a regex or rule-based extraction fallback for common failure modes.
4. For classification tasks, maintain a small deterministic ruleset for high-confidence cases and use the LLM only for ambiguous cases.
5. Log parse failures to identify when the LLM is drifting.
## When to Apply
- Any workflow where an LLM extracts or classifies data that downstream nodes consume.
- Any pipeline where consistency matters more than creative interpretation.
## Verification
- Test with malformed, ambiguous, and adversarial inputs.
- Confirm downstream nodes receive validated, structured data.
- Review parse-failure logs periodically.
## Related Patterns
- `ollama-structured-output-fallback`
- `json-escaping-downstream-api`