Dynamic Filters Deep-Dive

Standard RAG search performs only question → vector similarity search. Dynamic filters first narrow scope by metadata then perform vector search, improving accuracy.

Example

Question: “Show me revenue trends from the Finance team’s 2024 reports”

Method	Behavior	Result
Without filter	Vector similarity over all documents	Other teams’ docs may be mixed in
With filter	Narrow first by `department=Finance, year=2024`, then vector search	Only relevant docs retrieved

How to Configure Filters

Define filter fields

Click “Filter Settings” at the top right of the Knowledge Base detail screen, then ”+ Add Filter” at the bottom left of the modal to define a filter field.

Item	Description	Example
Field Name	Filter field name	”Department”, “Year”
Type	Data type	Single Select, Multi Select, Glossary, Date, Number
Mark as required	Warns files that leave this field empty when it’s set as required	Required ON/OFF toggle
Folder mapping	On folder upload, stores the folder name at this depth as the field value	None / Folder level 1, 2, 3
Filter Options	List of allowed values (Single Select / Multi Select types)	“Finance, HR, Engineering”
Agent Description	Description so the AI understands the filter	”Filter by the document’s department”
Extraction hint	Instruction used for AI auto-extraction	”Extract the department name from the filename”

Filters marked required show a * next to their name in the list on the left.

Choose extraction mode

Toggle Manual / AI at the top of the filter schema. Choosing AI reveals an AI model selector right below it. That model is shared by every filter’s AI extraction and AI description generation — it cannot be set per filter.

	Manual mode	AI mode
Input method	User enters per file	LLM auto-extracts from file content
Accuracy	100% (human-entered)	Depends on LLM performance
Time	Proportional to file count	One-click bulk extraction
Cost	None	LLM call cost
Recommended for	Few files or accuracy-critical	Many files needing fast classification

AI-mode results can be edited manually. The most efficient flow is bulk-extract with AI first, then fix only the errors.

Save

Click Save to save the filter schema. In AI mode, metadata is auto-extracted on subsequent file uploads.

Filter Type Details

Type	Search Behavior	Use Cases
Single Select	Exact match	Department, category, status
Multi Select	Match any one of multiple values	Tags, related teams
Glossary	Look up and search in a glossary	Product name
Date	Range search	Created date, expiry date
Number	Exact match	Year, version

Per-Type Limits

Each filter type has a fixed number of storage slots. Once they’re full, + Add Filter is disabled.

Type	Maximum
Single Select	4
Multi Select + Glossary (combined)	4
Date	2
Number	2

Multi Select and Glossary share the same slots. Create two Glossary filters and you can add at most two Multi Select filters.

Date Filter Input Format

Date filters accept varying precision:

Input	Meaning	Search Range
`2024`	All of 2024	2024-01-01 ~ 2024-12-31
`2024-03`	March 2024	2024-03-01 ~ 2024-03-31
`2024-03-15`	Specific date	That day only

Metadata State Display

At the far right of each file row, a filled fields / total filters badge is shown. The badge does not appear in a Knowledge Base with no filters defined.

Badge color	State	Condition	Tooltip on hover
🟢 Green	Complete	Every filter field has a value	Metadata complete (3/3)
🔵 Blue	Partial	All required fields are filled; some optional fields are empty	Metadata partially set (1/3)
🟠 Orange	Missing Required	At least one required field is empty	Required metadata missing (1/3)
⚪ Gray	Empty	Nothing is filled and no filter is marked required	No metadata extracted

While AI extraction is running, the file shows an “Extracting” badge with a spinner instead of the count.

The same 1/3 can appear in different colors. Color is decided by whether required fields are satisfied, not by how many fields are filled — if a required field is empty, the badge is orange no matter how much else you filled in.

Files in Orange (Missing Required) state may be omitted from filter searches. Make sure to fill in required fields.

AI Auto-Extraction

Writing Extraction Hints

The extraction hint is the instruction the AI uses when extracting metadata values from file content. Examples of good extraction hints:

Filter	Extraction hint
Department	”Extract the department from document content or filename. Options: Finance, HR, Engineering”
Year	”Extract the publication year from the document. If the filename contains a year, use it”
Document type	”Determine the document type. Options: policy, guide, report, form”

The AI analyzes the first ~4,000 characters of the file. The closer key information is to the beginning, the better extraction accuracy.

Running Extraction

Method	Where	When to Use
Auto-extraction	Runs automatically right after upload	Adding new files in AI mode
AI Extract (single)	Metadata panel that opens when you click a file	Re-extract one specific file
Extract (N) (bulk)	Toolbar above the file list	Extract every checked file at once
Clear metadata (N)	Toolbar above the file list	Clear filter values only; files and chunks stay

If you don’t see an extract button, your extraction hints are empty. The toolbar then shows only the notice “Configure filters to extract metadata”. The extract button appears once you’ve written at least one extraction hint in AI mode or linked a Glossary filter.

If already-extracted files are part of your selection, a dialog asks how to handle them. The default re-extracts every selected file, which also replaces values you edited by hand. Choose Extract only missing (N) to skip files that already have values.

Filling Metadata by Folder Upload

Set a filter’s Folder mapping and folder names become metadata values when you upload a whole folder. No LLM call is involved, so values are exact — the fastest path for documents you already sorted into folders. Depth counts the top folder you picked as level 0, and everything under it starts at level 1.

Project/21hg001/Contracts/a.pdf
  level 0   level 1   level 2

Map a “Project number” filter to Folder level 1 and the value becomes 21hg001
Map a “Document type” filter to Folder level 2 and the value becomes Contracts
Number types keep only the digits found in the folder name; Multi Select stores the folder name as a single value

How Agents Use Filters

When you connect a KB with dynamic filters to an agent, auto-filtering happens through this flow.

Step Details

Inform the AI about filters

On agent start, the system reads the KB’s filter schema and appends filter hints to the tool description.For example, with a “Department” filter, the AI knows “I can filter this KB by department”.

Extract filter values from the question

When a user asks a question, the AI auto-extracts filter values from the content.

Question: "Show me revenue trends from the Finance team's 2024 reports"

AI extraction:
  department → "Finance" (high confidence)
  year → 2024 (high confidence)

The AI only applies filters with high confidence. If filter values can’t be determined from the question, it falls back to standard vector search without filters.

Build the search filter

Extracted filter values are converted into a condition the search engine understands. With multiple values, only documents matching all of them (AND) survive.

Extraction: department=Finance, year=2024
     ↓
Condition: documents where department is "Finance" and year is 2024
     ↓
Passed to the search engine

Run filtered search

The search engine performs vector similarity search only on documents matching the filter conditions. Documents from other teams or years are excluded from search.

Generate the answer

Filtered relevant documents are passed to the AI to generate an accurate answer.

Tool Description vs. Per-Filter Agent Description

Two separate fields feed descriptions to the AI, and they do different jobs. The names look alike, so keep them apart.

	Tool Description	Agent Description (per filter)
Where	Top of the Knowledge Base detail screen	Filter Settings > each filter’s edit form
Scope	One per Knowledge Base	One per filter
What the AI decides with it	When to search this Knowledge Base	When to apply this filter

The label at the top of the screen is “Tool Description” — the agent calls a Knowledge Base like a tool. Its helper text reads “Describe when the agent should search this knowledge base.”

Leave the tool description empty and the KB’s general description is used instead.

	Without Tool Description	With Tool Description
KB selection	Decided by KB general description (may be inaccurate)	Clear purpose guidance for accurate selection
Filter usage	May not know filters exist	Knows which filter to use in which situation

AI auto-generation: Both fields have a generate button next to them — AI for the tool description, AI Generate for a filter’s agent description. The tool description is drafted from the KB name + description + file list + filter info, using whichever model is picked in the AI model selector under Filter Settings.

After changing the filter schema, regenerate the tool description. New filter info must be reflected there for the AI to use filters accurately.

Caveats

Are files with empty required fields omitted from search?

Yes. Files without values in fields marked Required are not included in results when searching with that filter condition. Look for orange badges and fill in values.

Can AI extraction be wrong?

AI extraction analyzes about the first 4,000 characters of the document — if key information is in later sections, extraction can be inaccurate. Review results and edit manually as needed.

What happens to existing metadata when I change the filter schema?

Adding a filter field leaves existing files’ values empty. Run bulk extraction in AI mode to populate the new field. Removing a field also removes that metadata.

Rather than adding too many filters to one Knowledge Base, consider splitting into separate Knowledge Bases if you need finer classification.

Workspace

AI Agents

Knowledge & Data

Safety

Customization

Example

How to Configure Filters

Define filter fields

Choose extraction mode

Save

Filter Type Details

Per-Type Limits

Date Filter Input Format

Metadata State Display

AI Auto-Extraction

Writing Extraction Hints

Running Extraction

Filling Metadata by Folder Upload

How Agents Use Filters

Step Details

Inform the AI about filters

Extract filter values from the question

Build the search filter

Run filtered search

Generate the answer

Tool Description vs. Per-Filter Agent Description

Caveats

​Example

​How to Configure Filters

Define filter fields

Choose extraction mode

Save

​Filter Type Details

​Per-Type Limits

​Date Filter Input Format

​Metadata State Display

​AI Auto-Extraction

​Writing Extraction Hints

​Running Extraction

​Filling Metadata by Folder Upload

​How Agents Use Filters

​Step Details

Inform the AI about filters

Extract filter values from the question

Build the search filter

Run filtered search

Generate the answer

​Tool Description vs. Per-Filter Agent Description

​Caveats

Example

How to Configure Filters

Filter Type Details

Per-Type Limits

Date Filter Input Format

Metadata State Display

AI Auto-Extraction

Writing Extraction Hints

Running Extraction

Filling Metadata by Folder Upload

How Agents Use Filters

Step Details

Tool Description vs. Per-Filter Agent Description

Caveats