Entity Extraction

Overview

Entity extraction is how the extension detects companies, investors, and people from page content. Each supported site has a SiteAdapter — a declarative module that knows how to parse entities from that site's DOM structure. The adapter runs in a content script, extracts entities, and sends them to the service worker for resolution against the DMTR intelligence database.

How it works

1. The content script (src/content/index.ts) checks the current URL against all registered adapters
2. The matching adapter's extractEntities(document) method runs, returning an array of ExtractedEntity objects
3. Each entity includes: name, type (startup/investor/fund/person), confidence (0–1), source (which selector matched), and optional metadata
4. The content script sends the entities to the background service worker via chrome.runtime.sendMessage
5. The side panel calls POST /v1/intelligence/resolve with the entity name to match it against the Blackmoat corpus

SiteAdapter interface

interface SiteAdapter {
  name: string;
  patterns: Array<{ hostname: string; pathname?: string }>;
  extractEntities(doc: Document): ExtractedEntity[];
  injectionPoints(doc: Document): InjectionPoint[];
}

interface ExtractedEntity {
  name: string;
  type?: "startup" | "investor" | "fund" | "person";
  confidence: number;
  source: string;
  metadata?: Record<string, unknown>;
}

Entity resolution

Extracted names are resolved to canonical Blackmoat entities via the /v1/intelligence/resolve endpoint. The resolver uses three strategies in order:

1. Alias lookup — exact match against a curated alias table (e.g., "a16z" → FC:ORG:a16z)
2. Slug/name match — exact match on entity slug or normalized display name
3. Fuzzy match — Levenshtein distance with configurable threshold (default 0.75)

Each match returns a confidence score. The side panel uses the highest-confidence match.

Confidence scoring

Current adapters

• PitchBook — extracts company name from profile headers, investor names from deal tables. See PitchBook.

Bloomberg, AlphaSense, Carta, and custom site adapters are planned but not yet implemented.