> For the complete documentation index, see [llms.txt](https://oodlz-ai-studio.gitbook.io/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://oodlz-ai-studio.gitbook.io/docs/inbound-agent/the-brain/four-specialist-tools.md).

# Four Specialist Tools

Four purpose-built tools — brands, FAQ, other, human escalation — each with tuned retrieval, called only when the intent demands.

> **One tool per intent — calibrated retrieval, no over-fetching, no LLM call where the answer is in memory.**

A chatbot with one knowledge base answers every question the same way. It cannot distinguish "how does cashback work" from "show me Nike deals" from "walk me through the app" from "I need to speak to someone". The Oodlz inbound agent has 4 purpose-built tools and routes to the right one based on classified intent.

## How It Works

The agent's system prompt declares 4 tool slots. The intent classifier and the agent's reasoning together pick which one fires per turn. Memory questions are answered directly from the conversation history — no tool call at all.

### `brands` — Live Cashback Catalogue

Vector search against the brand catalogue namespace. `topK` is set to 8 — enough candidates for the re-ranking layer to deduplicate, promote exact name matches, and surface featured brands and the highest cashback rates first. The catalogue is kept fresh by a separate upload pipeline.

### `faq` — General Platform Knowledge

Vector search against the FAQ namespace. `topK` is set to 1. Precision over volume — the single most relevant document is the answer. Used when the agent classifies the message as a "how does this work" question rather than a brand or product query.

### `other` — Tutorials And Walkthroughs

Vector search against the visual-and-instructional namespace. Browser extension setup with screenshots, step-by-step onboarding, multi-step how-to guides. Returns whatever the retrieval call surfaces — these queries need full sequential context, not a single hit.

### `human_escalate` — Human Handoff

HTTP tool, not a vector search. The agent collects and validates the user's email and issue description in conversation before the tool is called even once. The tool POSTs the structured payload — user email, issue description, session reference, original query — to the [human escalation pipeline](/docs/inbound-agent/the-knowledge/human-escalation.md). The agent confirms receipt to the user with a reference ID.

## What You Get

* Right tool per intent — brand queries hit the catalogue, FAQ queries hit precision retrieval, walkthroughs hit step-by-step content, escalation requests hit a human.
* No over-fetching — `topK` is tuned per namespace, so the agent doesn't sift through irrelevant results to find the one it needs.
* No tool call on memory questions — previous-turn references are answered directly from session history with no latency.
* Validated escalation — email and issue description are collected and confirmed before the human handoff fires, so support never sees a half-formed escalation.