kartika09/mas-reliability-capstone

How kartika09/mas-reliability-capstone works

This repository is a starter multi-agent research-analysis system. A researcher submits a question to the pipeline, which plans evidence and critical subtasks, optionally retrieves supporting abstracts, and has independent agents produce outputs. Depending on the experiment condition, the pipeline validates the evidence handoff, checks peer consistency, and may invoke a guard and recovery before the planner synthesizes a report. It records messages and run results for evaluation. Retrieval and fault injection are optional branches; several research mechanisms remain documented stubs. The dataset loader and logger details are inferred from documented interfaces, not sampled source.

Main entry points: pipeline.py Run pipeline · models.py Data contracts · retriever.py Embedding retrieval · agents.py Evidence agent

Components 19

Pipeline orchestration

Run pipeline pipeline.py
Fault injection fault_injection/injector.py
Research report

Research and retrieval

Planner agent agents/agents.py
Evidence agent agents/agents.py
Critical agent agents/agents.py
LLM client llm_client.py
SciFact dataset mas_scifact_dataset (1).json
Dataset loader dataset.py
Embedding retrieval retrieval/retriever.py
In-memory index retrieval/retriever.py

Validation and recovery

Sequential validator validation/sequential_validator.py
Peer consistency peer_consistency/checker.py
Guard decision guard/decision.py
Recovery guard/recovery.py

Contracts and evaluation

Data contracts schemas/models.py
Run logger evaluation/logger.py
Run result schemas/models.py

Other components

Researcher

Connections 30

Researcher → Run pipeline: submits question
Run pipeline → Planner agent: decomposes and synthesizes
Planner agent → LLM client: requests completion
Run pipeline → Evidence agent: requests evidence
Run pipeline → Critical agent: requests critique
Evidence agent → LLM client: requests completion
Critical agent → LLM client: requests completion
SciFact dataset → Dataset loader: loaded by
Dataset loader → Embedding retrieval: provides corpus
Embedding retrieval → In-memory index: builds index
Run pipeline → Embedding retrieval: retrieves passages
Embedding retrieval → Evidence agent: supplies context
Embedding retrieval → Critical agent: supplies context
Run pipeline → Fault injection: optionally injects fault
Run pipeline → Sequential validator: validates handoff
Run pipeline → Peer consistency: checks agreement
Run pipeline → Guard decision: requests decision
Run pipeline → Recovery: recovers flagged output
Run pipeline → Peer consistency: verifies recovery
Run pipeline → Sequential validator: verifies recovery
Run pipeline → Run logger: logs messages and run
Run pipeline → Research report: returns report
Run pipeline → Run result: creates run result
Run pipeline → Data contracts: uses contracts
Planner agent → Data contracts: parses and returns shapes
Evidence agent → Data contracts: parses output shape
Critical agent → Data contracts: parses output shape
Sequential validator → Data contracts: returns validation result
Peer consistency → Data contracts: returns peer result
Guard decision → Data contracts: returns guard decision

More Python diagrams: ghostfrankwu/sustech_tools · manish-9245/facial-emotion-recognition-using-opencv-and-deepface · nando0x/prospectos · prajwalvandana/maestro-cli · rawheel/fastapi-boilerplate

Drawn by GitDiagram on September 27, 2026 from the repository's file tree, README and source files. Read as Markdown ·