company-research

Company Research CLI

A one-shot command-line agent for company research over the Diffbot Knowledge Graph and the live web. Ask a question in plain English; the agent picks its own approach, may iterate, and cites the entity IDs / URLs it used. Useful for shell scripting or quick spot checks.

This is a TypeScript port of the company_research example from the Python langchain-diffbot package.

Prerequisites

• Node.js >= 20 and pnpm.
• DIFFBOT_API_TOKEN and ANTHROPIC_API_KEY in the environment. Copy ../.env.example to ../.env and fill them in (the CLI loads examples/.env).
• The parent package must be built: @diffbot/langchain's entry points resolve to dist/, which isn't produced until you build it.

Setup

From the repo root:

pnpm build

Then, from this directory:

cd examples/company-research
pnpm install

Run it

pnpm start "What companies in Austin work on robotics?"
pnpm start --quiet "Who are the executives at Diffbot?"
pnpm start "What did Diffbot announce most recently?"

By default the agent prints its tool calls and intermediate responses as it works. Pass --quiet to print only the final answer.

Example invocation

$ pnpm start "Who founded Diffbot and what does the company do?"

  ▸ calling: search_kg({"dqlQuery":"type:Organization name:\"Diffbot\""})
  ↳ tool search_kg: [{"summary":"Diffbot is a...","id":"<entity-id>","type":"Organization",...}]...
  ▸ calling: web_search({"query":"Diffbot founder"})
  ↳ tool web_search: [{"title":"Diffbot - Wikipedia","pageUrl":"https://en.wikipedia.org/wiki/Diffbot",...}]...

Diffbot was founded by Mike Tung (Diffbot, id=<entity-id>). The company
builds a Knowledge Graph of the public web using computer vision and
machine learning to extract structured data from web pages (diffbot.com).

This transcript is illustrative (shaped by hand to show the trace format), not a captured run — this environment has no DIFFBOT_API_TOKEN / ANTHROPIC_API_KEY, so the example has not been executed live. Run it yourself with real credentials to see actual output.

Tools

The agent has five tools backed by Diffbot, and chooses which to use per question:

• inspect_ontology(op, name?, search?) — look up the KG schema (entity types, a type's fields, taxonomy/enum values) so it writes DQL with real field paths.
• probe_dql(queries) — get hit counts for several DQL variants at once, to check a query is well-shaped before committing.
• search_kg(dqlQuery) — Knowledge Graph search via DQL.
• web_search(query) — natural-language web search, for current/news-y info or when the KG comes up short.
• extract_url(url) — fetch and read a single web page.

The intended loop is introspect (ontology) → probe → search → refine, with web_search + extract_url as the fallback when the KG doesn't have it.

Model

Defaults to anthropic:claude-haiku-4-5 because a multi-step agent loop on a fresh Anthropic account can blow past Sonnet's 30k input-tokens-per-minute Tier 1 limit. Override with any provider:model string createAgent understands:

COMPANY_RESEARCH_MODEL=anthropic:claude-sonnet-4-6 pnpm start "..."

Package classes exercised

From @diffbot/langchain:

• DiffbotKnowledgeGraphRetriever (via search_kg)
• DiffbotWebSearchRetriever (via web_search)
• DiffbotExtractTool (via extract_url)
• DiffbotOntologyTool (via inspect_ontology)
• DiffbotDQLProbeTool (via probe_dql)

All five share a single DiffbotClient from @diffbot/typescript, built once and reused for the whole process so the agent's tool loop shares one connection pool.

Layout

company-research/
├── src/
│   ├── agent.ts    # createAgent + system prompt + model selection
│   ├── tools.ts     # the five Diffbot-backed tools
│   └── cli.ts        # argv parsing, runs one question, renders the trace
├── package.json
├── tsconfig.json
└── README.md