SPB Git

spb/search-box Public

Agentic web research engine — hypotheses, verbatim evidence, contradictions, sourced answers streamed live. Claude Opus 5 + Firecrawl + PostgreSQL.

airesearchagentsclaudenextjstypescript
TypeScript 76.9% CSS 18.7% SQL 2.1% JavaScript 1.8% Shell 0.5%
1.6 KB
View source Raw Blame History

# ResearchState

All research state lives in PostgreSQL (packages/db/migrations/001_init.sql), not in the model's context window. Sessions are resumable and replayable.

# Entities

  • Session — question, status (pending → running → synthesizing → completed|failed), objectives, budgets, final answer.
  • Source — every URL seen. found → fetched|failed. Fetched sources store up to 120k chars of markdown. citation_index is assigned mechanically at synthesis.
  • Evidence — verbatim quote from a source, attached to a claim with a stance (supports | contradicts | context). Quotes are soft-verified against stored content; unverified quotes get flagged in note.
  • Claim — falsifiable statement with status (exploring | supported | contradicted | uncertain) and confidence (0..1 probability).
  • Contradiction — first-class record of credible disagreement, referencing the conflicting evidence ids.

# Provenance chain

answer sentence → [n] marker → citation_index → source → evidence quotes → claims

Citation indices are derived mechanically from state (sources ordered by first evidence use). The synthesis model never invents numbering.

# Events

Every mutation appends exactly one event row (research_events, monotonic seq). The event log is the replay/debug record: replaying a session's events from seq 0 reconstructs everything the user saw live.