Skip to content

ADR.md

Latest commit

 

History

History
209 lines (65 loc) · 1.77 KB

ADR.md

File metadata and controls

209 lines (65 loc) · 1.77 KB

Architecture Decision Record (ADR)

ADR-001: MVP Architecture & Tooling

Status

Accepted (MVP)

Context

The system requires:

  • Explicit reasoning steps

  • Conditional refusal or uncertainty

  • Evidence-grounded outputs

  • Fast prototyping with minimal frontend work

Decisions

1. LangGraph for Orchestration

LangGraph will define a deterministic, node-based workflow.

Proposed Nodes

  1. Claim Extraction

  2. Search & Retrieval (Tavily)

  3. Evidence Filtering

  4. Evidence–Claim Comparison

  5. Verdict Assembly

Rationale

  • Clear state transitions

  • Easy to reason about failures

  • Avoids opaque agent loops

2. OpenAI API for Language Reasoning

Used for:

  • Sub-claim extraction

  • Evidence relevance assessment

  • Verdict reasoning

Rationale

  • Strong instruction adherence

  • Better grounding behavior for evidence-based tasks

  • Lower implementation risk for MVP

3. Tavily for Web Search

Used for retrieving candidate sources and original reporting.

Rationale

  • Search-focused API

  • Structured responses

  • Faster than custom scraping

4. Gradio for UI

Gradio will serve as the presentation layer.

Rationale

  • Minimal frontend complexity

  • Rapid iteration

  • Suitable for demos and portfolio review

  • Clear separation from backend logic

Alternatives Rejected

  • Multi-agent debate systems (scope risk)

  • React/Next.js frontend (unnecessary complexity)

  • Autonomous browsing agents (instability, complexity)

Consequences

Positive

  • Inspectable reasoning

  • Easy to explain in interviews

  • Small, coherent codebase

Negative

  • Not production-grade

  • Limited scalability

  • Dependent on external APIs