> ## Documentation Index
> Fetch the complete documentation index at: https://docs.rippit.com/llms.txt
> Use this file to discover all available pages before exploring further.

# External research

> Search the web, find people, and ask AI assistants a question, for every row of a table or on a schedule, and save each finding as a row.

Your data stops at what customers told you. It doesn't say which account just announced layoffs, who leads support at a prospect, or whether ChatGPT recommends you. Research brings that outside information into Rippit: web pages, people, and how AI models such as ChatGPT answer a question. In a [pipeline](https://docs.rippit.com/pipelines/overview.md), research runs for every row of a table or on a schedule, and each finding is saved as a row you can query, chart, and act on. In chat, the same searches answer a question once.

## The three search methods

| Method | What it does | What comes back | Providers |
| - | - | - | - |
| Web search | Finds public pages that match what you describe, and reads pages at URLs you already have | Title, link, publish date, the relevant passages, and the page text | Exa by default, or Parallel, Ceramic, or DeepSeek |
| People search | Finds public professional profiles by name, role, company, expertise, or location, and can look up work email and phone number | Name, role, company, location, profile link, and any contact details you asked for | Firecrawl by default, or Clay or Exa |
| Agent search | Asks AI models the same question with their own web search turned on, and records how each one answers | The answer, its citations, and the searches the model ran | ChatGPT, Claude, and Gemini |

Use web search to track pages and news, people search to find prospects or contacts, and agent search to see how AI models talk about a topic, a brand, or a competitor.

## Research in a pipeline

A pipeline repeats the research and keeps the results:

* **For each item in a table.** Research each account, competitor, person, or question. Rippit rolls the table up to one row per item, and the research runs once for each row.
* **On a schedule.** Run the same research every day, week, or month, and build a history you can compare over time.
* **Into a table.** Each finding is saved as a row, so you can query it, chart it, and ask about it in chat.
* **As one step among others.** Later steps can classify each finding with an [AI column](https://docs.rippit.com/pipelines/steps.md#ai-columns), roll findings up by account, and [send the result](https://docs.rippit.com/pipelines/steps.md#actions) to Slack or email.

Describe the research and how often it should run:

```text wrap theme={null}
Every Monday, check each company in my competitors table for pricing or product changes on its website, and save what changed with a link.
```

```text wrap theme={null}
Every week, ask ChatGPT, Claude, and Gemini each question in my buyer questions table. Save every answer with its citations, and add a column for whether we were mentioned.
```

```text wrap theme={null}
For each company in my target accounts table, find the head of support and look up their work email.
```

In the pipeline, research is an [agent run](https://docs.rippit.com/pipelines/steps.md#agent-runs) that has been given search tools. Agent runs work on the rows of a rollup, so the research runs once for each row the rollup produces. A simple web or people lookup can also run without an agent, as a read step of an [action](https://docs.rippit.com/pipelines/steps.md#actions). Agent search always runs inside an agent run, which waits for every model to answer before it writes the row.

Work-email and phone lookups run only when you ask for them. When a row is retried, agent search reuses the search it already paid for, while web and people searches run again. Workspace access, configured providers, model permissions, and spending limits still apply. [Tools an assistant uses](https://docs.rippit.com/pipelines/reference.md#tools-an-assistant-uses) lists the search tools by name.

## Research in chat

Ask in plain language and Rippit runs the search in the conversation, cites what it used, and answers.

* "Find recent reviews of AI support tools and summarize what buyers complain about."
* "Who leads customer support at Acme? Include their profile link."
* "Ask AI assistants: Which support QA tools do you recommend?"

For the last one, Rippit sends the exact question independently to up to three models with their own web search enabled. The default comparison uses one current model each from Google, Anthropic, and OpenAI. You can name one to three models yourself. When a default model is unavailable, chat shows which model took its place.

Open the **Agent search** card in the conversation, then select a model to read its full answer, citations, and reported search activity. The card stays visible when the chat's execution steps are collapsed, and you can reopen it after reloading the chat.

## Choose a provider

Every provider is built into Rippit, so there is no account or API key to set up. Name the one you want when you describe a pipeline or ask in chat. Otherwise Rippit uses the default.

### Web search

| Provider | Pick it when |
| - | - |
| Exa (default) | Most research. It searches by meaning rather than exact keywords and returns the most complete page content, with publish dates |
| Parallel | You want to compare coverage. Similar quality, with page excerpts and publish dates |
| Ceramic | Cost matters most. Much cheaper keyword search, with snippets only and no publish dates |
| DeepSeek | You want a cheaper option that still returns page content and publish dates |

A search can be limited to certain sites, exclude sites, or keep only pages published in a date range. Date ranges need a provider that returns publish dates.

### People search

| Provider | Pick it when |
| - | - |
| Firecrawl (default) | Most people research. It finds people who match and looks up work email and phone number in the same search |
| Clay | Firecrawl's results come back thin for your audience. It searches Clay's people data and also looks up work email and phone number |
| Exa | You want a list of people in a role or industry and don't need contact details. If you ask for them, Rippit looks them up through Firecrawl or Clay |

Contact lookups are off by default. Ask for work email, phone number, or both; asking for "people," "leads," or "contacts" alone does not turn them on. A person stays in the results even when a lookup comes back empty, and each contact detail carries a status, so you can tell one that was not found from one that failed. Rippit reports a detail as verified only when the provider does.

Clay here is the built-in search provider. To work with your own Clay workspace, including its functions, audiences, and credits, connect the [Clay MCP server](https://docs.rippit.com/agentic-chat/connectors/clay.md).

### Agent search

Agent search covers models from OpenAI (ChatGPT), Anthropic (Claude), and Google (Gemini). By default it compares one model from each, and you can name one to three yourself.

* **The question goes to each model as-is**, with its own web search on and no extra instructions from Rippit, so the answer reflects what the model says on its own.
* **These are API calls, not a recording of the apps.** Answers can differ from what someone sees in ChatGPT, Claude, Gemini, or Google's AI Overviews.
* **A model can answer without searching**, and providers differ in which search details they expose.
* **A model that fails is reported, not dropped**, so a comparison always shows what happened to each one.
* **Gemini cites sources through Google redirect links.** Rippit keeps those links as Google returns them.

## Ask for specific fields

Any search can return typed fields alongside its results. Ask for concrete fields, such as a published price and currency, or a person's current role and employer, and say what type each is: text, a list of text, an integer, a decimal, or true or false. Rippit extracts each field from that result's evidence.

Each field reports whether it was extracted, was unavailable in the evidence, or failed. Missing evidence produces an unknown value rather than a guess, and unknown is different from false, zero, or an empty list. The original answer or source page stays attached so you can inspect the evidence. Typed fields make results easier to use; they do not independently verify a source's claims.

Extraction adds AI cost and latency. Agent search extracts from each model's original answer without changing the question it was asked.

## Cost

Research uses AI credits for the model doing the research, plus provider costs for the search tools. Contact lookups use extra provider credits, and Rippit reports the credits each one used. In a pipeline, every run pays for its searches again, so match the schedule to how often the answer really changes.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.