How kaymansrinivasan/air-gapped-rca works
airgap-rca is a prototype for grounded root-cause analysis of wafer-sort failures on offline edge accelerators. Its intended engineer workflow starts with immutable tester logs, validates and chunks them with source-line provenance, retrieves similar historical evidence, drafts ranked possible causes with citations, and checks claims before answering or refusing. The sampled code implements synthetic data generation, validation, chunking, Chroma ingestion, case evidence retrieval, and local draft generation; the README describes verification as part of the proposed workflow. Some intended stages, including keyword search, are not represented by sampled implementation files.
Main entry points: retrieve_case_evidence.py Case Evidence Retrieval · chunk.py Source-Line Chunker · src/generate.py Local Draft Generator · product_a_v1/generate.py Wafer Log Generator
Components 17
Synthetic Data
- Wafer Log Generator
syn_data/product_a_v1/generate.py - Wafer Sort Logs
syn_data/product_a_v1/logs - Dataset Validator
syn_data/product_a_v1/validate.py - Case Histories
syn_data/product_a_scenarios_v1 - Scenario Builder
scripts/create_product_a_scenarios.py
Evidence Preparation
- Source-Line Chunker
src/chunk.py - Observation Chunks
- Vector Ingestion
src/chroma_ingest.py - Local Chroma Store
Case Retrieval
- Failure Query
syn_data/product_a_scenarios_v1/case_07/observation.json - Case Evidence Retrieval
src/retrieve_case_evidence.py - Evidence Bundle
Grounded Answer
- Local Draft Generator
src/generate.py - Local Ollama
- Citation Verification
src/verify.py - Supported Answer / Refusal
Other components
- Engineer
Connections 18
- Wafer Log Generator → Wafer Sort Logs: writes
syn_data/product_a_v1/generate.py - Wafer Log Generator → Dataset Validator: validates output
syn_data/product_a_v1/generate.py - Source-Line Chunker → Dataset Validator: independently validates
src/chunk.py - Source-Line Chunker → Wafer Sort Logs: reads source
src/chunk.py - Source-Line Chunker → Observation Chunks: writes chunks
src/chunk.py - Scenario Builder → Observation Chunks: reads observations
scripts/create_product_a_scenarios.py - Scenario Builder → Case Histories: writes records
scripts/create_product_a_scenarios.py - Vector Ingestion → Observation Chunks: reads chunks
src/chroma_ingest.py - Vector Ingestion → Local Chroma Store: upserts vectors
src/chroma_ingest.py - Engineer → Failure Query: supplies failure
README.md - Case Evidence Retrieval → Failure Query: loads query
src/retrieve_case_evidence.py - Case Evidence Retrieval → Case Histories: loads histories
src/retrieve_case_evidence.py - Case Evidence Retrieval → Local Chroma Store: queries candidates
src/retrieve_case_evidence.py - Case Evidence Retrieval → Evidence Bundle: creates evidence
- Local Draft Generator → Evidence Bundle: reads evidence
src/generate.py - Local Draft Generator → Local Ollama: requests draft
src/generate.py - Local Draft Generator → Citation Verification: submits claims
README.md - Citation Verification → Supported Answer / Refusal: checks answer
README.md