Skip to main content

Advanced Search

ZenSearch provides powerful search capabilities combining semantic understanding with precise filtering. Master these features to find exactly what you need.

Search Architecture​

ZenSearch uses a hybrid approach combining:

Dense Embeddings​

  • Semantic understanding: Finds conceptually similar content
  • Meaning over keywords: "car" matches "automobile"
  • Context awareness: Understands intent behind queries

Sparse Embeddings​

  • Keyword precision: Exact term matching
  • Technical terms: Catches specific jargon
  • Names and codes: Finds exact identifiers

Fusion​

Results from both methods are combined using sophisticated ranking algorithms to provide the best of both worlds.

Retrieval Pipeline​

Query
  ↓
Intent Classification
  ↓
Query Expansion (optional)
  ↓
Permission Filtering
  ↓
Hybrid Search (Dense + Sparse)
  ↓
Context Decay (time-based weighting)
  ↓
Faceted Filtering
  ↓
Cross-Encoder Reranking
  ↓
Citation Grounding
  ↓
Context Enrichment
  ↓
Results

Query Types​

Natural Language Questions​

Ask questions as you would to a colleague:

"What is our policy on remote work?"
"How do I configure the database connection?"
"Who is responsible for the Q4 budget?"

Keyword Searches​

Use specific terms for precision:

"API authentication token"
"error code 5001"
"employee handbook 2024"

Combined Queries​

Mix natural language with specific terms:

"How do I fix error ERR_CONNECTION_REFUSED?"
"What are the steps to deploy version 2.3.1?"

Query Expansion​

ZenSearch can automatically expand your query to improve results:

How It Works​

  1. Your query is analyzed
  2. Alternative phrasings are generated
  3. Multiple searches run in parallel
  4. Results are merged and deduplicated

Example​

Original: "How to fix login issues"

Expanded:
- "login issues troubleshooting"
- "authentication problems resolution"
- "sign in errors fix"
- "login failure solutions"

Available Facets​

FacetDescriptionExample Values
TopicsContent categoriesTechnology, Finance, HR
DepartmentsOrganizational unitsEngineering, Sales, Marketing
LanguagesDocument languageEnglish, Spanish, French
SentimentsContent tonePositive, Neutral, Negative
Date RangeCreation/modificationLast 7/30/90 days, Custom

Using Facets​

  1. Perform a search
  2. View facets in the sidebar
  3. Click to filter by facet values
  4. Combine multiple facets
  5. Clear filters to broaden results

Dynamic Facets​

Facets update based on current results:

  • Counts reflect filtered results
  • Unavailable facets are hidden
  • Values sorted by relevance

Context Decay (Time-Based Weighting)​

ZenSearch applies time-based score weighting so that more recent documents rank higher than stale ones, all else being equal. This ensures your search results reflect current information rather than outdated content.

How It Works​

Each document's relevance score is adjusted by a decay factor based on the document's age. The decay follows a half-life model: after one half-life period, a document's time-based weight drops to 50%.

Configuration​

Context decay is configurable per collection:

SettingDescriptionDefault
EnabledWhether time decay is activefalse
Half-lifeDuration after which weight drops to 50%180 days

Use Cases​

  • News or announcements: Short half-life (7-14 days) to surface the latest updates
  • Policies and procedures: Longer half-life (180+ days) since these change infrequently
  • Code documentation: Medium half-life (60-90 days) to balance freshness with stability
info

Context decay adjusts ranking, not visibility. Older documents still appear in results if they are highly relevant — they simply rank lower than equally relevant newer documents.

Multi-Modal Search (Image Content)​

ZenSearch extracts and indexes images embedded in documents, making visual content searchable alongside text. When a document contains images (diagrams, charts, screenshots, photos), each image is processed through a vision model that generates a natural-language description.

How It Works​

  1. During document parsing, images are extracted as IMAGE structural units
  2. A vision model analyzes each image and produces a text description
  3. The description is embedded and indexed alongside the document's text content
  4. When you search, image descriptions are included in the retrieval pipeline

What This Means for You​

  • Search for "architecture diagram" and find images matching that description
  • Charts and graphs are described with their data points and trends
  • Screenshots of UIs are described with their visible elements
  • All image content participates in hybrid search (dense + sparse)

Supported Formats​

Images in the following document types are automatically extracted:

  • PDF files (embedded images and figures)
  • Word documents (.docx)
  • PowerPoint presentations (.pptx)

Citation Grounding​

After generating a response, ZenSearch verifies that each citation correctly attributes to its claimed source. This post-synthesis verification step catches misattributed citations and automatically corrects them.

How It Works​

  1. The AI generates a response with inline citations [1], [2], etc.
  2. A grounding check compares each cited claim against the actual source content
  3. Misattributed citations are corrected or removed
  4. The final response contains only verified source attributions

Benefits​

  • Higher trust in cited sources
  • Reduced hallucination in attribution
  • Each citation points to content that genuinely supports the claim

Progressive Retrieval​

When the retrieval pipeline detects that initial search results have low confidence or insufficient coverage for a query, it autonomously fetches additional context. The system re-queries with refined terms, broader scope, or alternative phrasings to improve answer quality.

This is particularly useful for:

  • Niche or highly specific queries where the first pass returns few results
  • Questions that span multiple topics or collections
  • Queries where the answer requires synthesizing information from several documents

Progressive retrieval is transparent — you may see a brief "Searching for more context..." indicator while additional sources are being gathered.

Cross-Encoder Reranking​

What Is Reranking?​

After initial retrieval, a cross-encoder model reranks results for better precision:

  1. Initial retrieval: Fast, broad search
  2. Reranking: Deep analysis of top candidates
  3. Final order: Most relevant results first

Benefits​

  • More accurate relevance scores
  • Better handling of complex queries
  • Improved result ordering

Coverage Information​

Understanding Coverage​

Search results include coverage metrics showing completeness:

  • Full coverage: All relevant content found
  • Partial coverage: Some content may be missing
  • Warnings: Potential gaps in results

Coverage Indicators​

Results: 15 documents found
Coverage: 94% (3 semantic units pending indexing)

⚠️ Some content from GitHub connector is still syncing

Search Modes​

Chat Mode​

Best for:

  • Questions needing synthesized answers
  • Multi-turn conversations
  • Research and exploration

Features:

  • AI-generated responses
  • Source citations
  • Follow-up capability

Search Mode​

Best for:

  • Finding specific documents
  • Browsing available content
  • Detailed filtering

Features:

  • Document list results
  • Faceted filtering
  • Preview snippets

Scope and Collections​

Collection Scoping​

Control search boundaries:

ScopeUse Case
All CollectionsCompany-wide search
Single CollectionDepartment-specific search
Multiple CollectionsCross-functional research

Setting Scope​

  1. Click the Scope dropdown
  2. Select collections to include
  3. View document counts
  4. Search within selection

Answer Shape​

Query Classification​

ZenSearch classifies queries to optimize responses:

ShapeDescriptionExample
EnumerativeList of items"What tools do we use?"
ProceduralStep-by-step"How do I submit expenses?"
ExploratoryOpen-ended"Tell me about our products"
ComparativeComparison"Compare Plan A vs Plan B"

Response Formatting​

Responses are formatted based on query shape:

  • Enumerative: Bulleted lists
  • Procedural: Numbered steps
  • Exploratory: Comprehensive overview
  • Comparative: Tables and comparisons

Meta-Questions​

About Your Knowledge Base​

Ask meta-questions about your indexed content:

"What topics are covered in our documentation?"
"Give me an overview of the engineering wiki"
"What data sources are connected?"
"Show me statistics about our content"

Meta-Question Indicators​

Meta-questions are indicated with badges:

  • Overview
  • Topics
  • Data Sources
  • Statistics
  • Capabilities

Search Tips​

Effective Queries​

StrategyExample
Be specific"Q4 2024 sales report" vs "sales"
Add context"Python API authentication" vs "authentication"
Use timeframes"last quarter", "2024", "recent"
Name specificsInclude project, team, or person names

Refining Results​

  1. Start broad, then narrow with facets
  2. Try alternative phrasings
  3. Use both chat and search modes
  4. Check suggested related queries

Interpreting Results​

IndicatorMeaning
High relevanceStrong match to query
Multiple citationsSynthesized from several sources
Recent dateCurrent information
Verified sourceFrom authoritative connector

Permissions and Access​

Search-Time Filtering​

ZenSearch enforces permissions at search time:

  1. Query is received
  2. User's access rights are checked
  3. Only accessible documents are searched
  4. Results exclude unauthorized content

Permission Types​

TypeDescription
UserIndividual access rights
GroupTeam or group membership
TeamWorkspace access
DomainOrganization-wide
PublicNo restrictions

Performance​

Speed Optimization​

ZenSearch optimizes for fast results:

  • Cached embeddings
  • Indexed metadata
  • Parallel searches
  • Incremental updates

Large Result Sets​

For queries with many results:

  • Pagination available
  • "Load more" functionality
  • Result count displayed
  • Coverage information shown

Troubleshooting​

No Results​

  1. Check collection scope
  2. Broaden search terms
  3. Remove filters
  4. Verify content is indexed

Irrelevant Results​

  1. Add more specific terms
  2. Use facet filters
  3. Try different phrasing
  4. Check query intent

Slow Searches​

  1. Narrow collection scope
  2. Simplify complex queries
  3. Check for large pending syncs
  4. Use specific filters

Next Steps​