Skip to content

PRD.md

Latest commit

 

History

History
135 lines (88 loc) · 3.6 KB

PRD.md

File metadata and controls

135 lines (88 loc) · 3.6 KB

Chase the Source — Product Requirements Document (PRD)

1. Problem Statement

Articles and social posts often contain interpretive or opinionated language that embeds implicit factual sub-claims. Readers cannot easily trace these claims back to original reporting or primary sources, nor determine whether the source material directly states, paraphrases, or contradicts the claim.

2. Goal

Given a short input text, the system will:

  1. Extract one factual sub-claim (even if embedded in opinionated language)
  2. Retrieve relevant sources via web search
  3. Classify the source-to-claim relationship: Direct, Paraphrase, Contradiction, or Not Found
  4. Present verbatim evidence showing where the claim originated

This is an ephemeral, single-session MVP focused on source attribution and transparency.

3. Non-Goals

  • No persistent storage or user accounts
  • No batch or multi-claim processing
  • No automated truth scoring
  • No political stance labeling
  • No browser extension or API productization

4. Target User

  • Journalists / researchers (prototype-level)
  • Technical evaluators (hiring managers)
  • Advanced users interested in source attribution

5. Core User Flow

  1. User pastes a paragraph (article excerpt, post, commentary)

  2. System extracts one factual sub-claim

  3. System retrieves sources using Tavily

  4. System evaluates evidence using grounded LLM reasoning

  5. System displays:

    • Extracted claim
    • Verdict
    • Sources and quoted evidence

6. Functional Requirements

6.1 Input

  • Free-form text (≤ 2 paragraphs)

6.2 Claim Extraction

  • System attempts to extract a factual sub-claim even if the input is opinionated
  • If no factual claim can be reasonably isolated, the system must state this explicitly

6.3 Source Retrieval

  • Original reporting is allowed

  • Sources may be:

    • Primary (transcripts, filings, datasets, direct statements)
    • Original reporting (first-party journalism)
    • Secondary reporting (clearly labeled)

6.4 Source Attribution Categories

  • Direct – Source states the claim verbatim or near-verbatim
  • Paraphrase – Source conveys the same meaning in different words
  • Contradiction – Source states the opposite or conflicts with the claim
  • Not Found – No sources found that address this specific claim

7. Output Requirements

7.1 Required Outputs

  • Extracted Claim (single sentence)

  • Attribution: Direct | Paraphrase | Contradiction | Not Found

  • Evidence List (1–5 items):

    • Source URL
    • Source type (primary / original reporting / secondary)
    • Verbatim quoted excerpt
    • How the source relates to the claim (direct quote, paraphrase, or contradiction)

7.2 Transparency

  • All attributions must be traceable to verbatim quoted text
  • System must explicitly note when sources are secondary (not original reporting)
  • System must show the "chain" from claim → source quote

8. UI Requirements (Gradio)

  • Single-page interface

  • Text input field

  • "Chase the Source" action button

  • Structured output sections:

    • Extracted Claim
    • Verdict
    • Sources & Evidence

No advanced settings or debug UI in MVP.

9. Success Criteria (MVP)

  • End-to-end execution < 30 seconds
  • Claim extraction is understandable and defensible
  • Attribution category matches the quoted evidence
  • System returns Not Found rather than fabricating a source relationship

10. Known Limitations

  • Temporal mismatches (old vs recent reporting)
  • Search result noise
  • Ambiguous or compound claims
  • No human-in-the-loop correction