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

# Targeted Discovery

> Active, account-first signal scanning powered by Exa AI.

<Warning>
  **Admin Only Feature:** Targeted Discovery is a high-power, credit-intensive pipeline currently gated behind a feature flag and requires the `signals:create` RBAC permission.
</Warning>

Unlike the continuous, passive web monitoring of the standard Market Radar, **Targeted Discovery** is an *active, account-first* scanning pipeline. It iterates over your prioritized accounts and actively hunts for specific signals on the live web.

## The Architecture

Targeted Discovery runs on a highly resilient, asynchronous **QStash worker pipeline**. This allows it to handle massive batches of accounts, respect provider rate limits, and resume gracefully if interrupted.

### Progressive Exa AI Query Packs

Instead of running one generic search, the engine uses 3 tiers of Progressive Query Packs. If Tier 1 (high precision, low recall) fails to find a signal, it gracefully falls back to Tier 2, and then Tier 3, ensuring maximum coverage while optimizing token spend. It actively searches across 9 core signal families.

### 4-Dimension Candidate Evaluation

When Exa AI returns a candidate URL, it isn't immediately accepted. The LLM Evaluator scores the candidate across 4 dimensions:

1. **Name Match:** Does the text actually refer to the target company?
2. **Domain Match:** Does the context align with the company's domain?
3. **Context Consistency:** Is the event described logically consistent?
4. **Signal Quality:** Is this a high-value signal or just SEO spam?

A candidate must score a minimum of **60/100** to proceed.

### 4-Layer Deduplication

To ensure you never see the same signal twice, candidates pass through a 4-layer dedup firewall before insertion:

1. Exact URL Match
2. Semantic Title Similarity
3. Payload SHA-256 Fingerprinting
4. Recent Time-Window Exclusion (prevents logging the same event multiple times in a week)