Skip to main content

How to Select the Right Focus in Chat

To get accurate and useful answers in NEXT AI Chat, you need to ensure that the system retrieves the right highlights.

Highlights act as the database for Chat. The quality of answers depends directly on how well this data is selected.

How Data Selection Works​

NEXT AI combines two complementary approaches:

  1. Semantic search (vector-based) → core retrieval layer
  2. Tag-based filtering (structured data) → precision & scaling layer

You do not need to manually choose between them.

Instead, NEXT automatically:

  • Uses tag-based filtering where relevant
  • Applies semantic search on top
  • Falls back to pure semantic search if no structure is available

Unlike traditional RAG systems, NEXT natively combines structured filtering and semantic retrieval, so you can start simple and improve over time without changing how you query.

1. Semantic Search (Core Layer)​

Semantic search is the primary retrieval mechanism in Chat. It finds highlights based on meaning, not exact wording.

Example:

Asking about "onboarding issues” will also retrieve highlights mentioning "getting started”, "setup”, or similar concepts.

When semantic search shines​

Semantic search works best when:

  • You are starting out with little structure
  • You want broad exploration
  • Language varies across sources
  • Concepts are fuzzy or loosely defined

Advantages​

  • No setup required
  • Captures intent, not just keywords
  • Works across all data sources

Limitations​

Semantic search operates on a limited candidate set of highlights.

This means:

  • Not all data is evaluated at once
  • Some relevant highlights may be missed at scale
  • Precision decreases for very specific queries

2. Tag-Based Filtering (Precision & Scale)​

Tag-based filtering uses explicit structure such as:

  • Tags (e.g. "Onboarding”, "Bug”)
  • Time (e.g. last 30 days)
  • Metadata (e.g. country, source)

Example:

"All highlights from the last month with the tag ‘Onboarding'”

When tag-based filtering shines​

Use structured filtering when:

  • You need precise control
  • You want full dataset coverage
  • You rely on explicit identifiers
  • Your dataset is large

Advantages​

  • High precision
  • Works across the full dataset
  • Predictable and controllable

Limitations​

  • Requires setup
  • Depends on tag quality
  • Needs maintenance over time

How NEXT Combines Both​

NEXT automatically blends both approaches:

  1. Apply tag-based filtering (if relevant tags exist)
  2. Run semantic search within the filtered set

If no useful tags are available, it defaults to semantic search.

This means:

  • You can start without configuration
  • You can gradually add structure
  • You never lose flexibility

What You Actually Do​

You don't configure retrieval directly. Instead, you improve it indirectly by adding structure:

Choosing the Right Strategy​

  • Don't over-engineer upfront
  • Let semantic search handle retrieval
  • Explore your data first

Add Structure Where It Helps​

Introduce tags when:

  • Results become too broad
  • You need segmentation (e.g. country, persona)
  • You want consistent slicing of data

When You Need More Than Semantics​

Some concepts are not explicitly mentioned in data:

"What do we know about recurring customers?”

This requires interpreting behavior and intent, not just matching text.

Semantic search alone may not capture this reliably.

Solution: Add Structure via Tagging Jobs​

Tagging Jobs allow you to:

  • Encode implicit concepts
  • Add consistent labels across all data
  • Improve retrieval quality for recurring questions

When Filtering Becomes Essential​

Use tag-based filtering when:

  • You need exact matches (e.g. competitor names)
  • You want complete coverage of data
  • You rely on clear attributes

Examples:

  • "Mentions of competitor X”
  • "Feedback from Germany in last 3 months”

Avoid Poor Tags​

Tags can reduce quality if misused.

Avoid tags that:

  • Exist only in part of your data
  • Are inconsistently applied
  • Represent vague or overlapping concepts

Example:

If "Onboarding” exists only in surveys, filtering by it removes interview data, leading to biased answers.

Rule of Thumb​

If you're unsure:

  • Start without filters → rely on semantic search
  • Add filters only when results are too broad or noisy
  • Avoid over-filtering early → it can hide relevant data
  • Avoid adding too many tags too early → structure should follow real usage

NEXT will often suggest ways to narrow down results automatically, so you don't need to get it right upfront.

Scaling Considerations​

  • Semantic search = flexible, but operates on a limited subset
  • Tag-based filtering = enables full dataset coverage

As your data grows:

  • Relying only on semantic search becomes insufficient
  • Adding structure becomes necessary for consistency and completeness

Improving Over Time​

Data selection evolves with usage.

Step 1: Observe​

  • Which questions work well?
  • Where do results feel incomplete?

Step 2: Identify Gaps​

Typical issues:

  • Missing structure
  • Too much noise
  • Biased subsets

Step 3: Improve​

  • Add or refine tags
  • Introduce Tagging Jobs
  • Improve consistency

Key Takeaway​

You don't need to choose a strategy upfront.

  • Semantic search is the foundation
  • Tag-based filtering improves precision and scale

Start simple, let NEXT guide you, and add structure only where it creates real value.