← Back to Home

Search Query Guide

Learn how to search the Monta Knowledgebase effectively using the query tool.

Search Modes

The knowledgebase supports three search modes to match different use cases:

Default Mode (Hybrid Search)

Best for: Most searches, natural language questions

Combines keyword matching with semantic understanding to find the most relevant results.

Example:

Query: "How do I set up a charging station?"
Mode: default

This finds documents that:

BM25 Mode (Keyword Search)

Best for: Exact phrases, technical terms, specific keywords

Finds documents containing your exact search terms using BM25 ranking.

Example:

Query: "OCPP 2.0.1 configuration"
Mode: bm25

Use when you know the exact terminology or need precise matches.

Vector Mode (Semantic Search)

Best for: Conceptual queries, finding related content

Finds documents with similar meaning, even if they don't use your exact words.

Example:

Query: "problems with billing customers"
Mode: vector

Might find documents about "invoice errors", "payment issues", or "charging session costs" even without those exact terms.

How Vector Search Works

Vector search uses embeddings and chunking to understand the semantic meaning of your query and find relevant content.

The Process:

  1. Document Chunking (at index time)

    • Large documents (>3KB) are split into chunks of 3000 characters
    • Each chunk overlaps by 300 characters to maintain context continuity
    • Chunks break at sentence boundaries when possible for readability
    • Each chunk maintains a reference to its parent document

  2. Embedding Generation (at index time)

    • Every chunk is converted into a 1536-dimensional vector using OpenAI's text-embedding-3-small model
    • This vector captures the semantic meaning of the text
    • Full documents are also stored with their complete content alongside embeddings

  3. Query Embedding (at search time)

    • Your search query is converted into the same 1536-dimensional vector space
    • This allows mathematical comparison of semantic similarity

  4. Chunk Search (RAG - Retrieval Augmented Generation)

    • The system searches through all document chunks using k-Nearest Neighbors (kNN)
    • Finds chunks with vectors most similar to your query vector (cosine similarity)
    • Multiple chunks from the same document may match your query

  5. Document Retrieval

    • Matching chunks are grouped by their parent document ID
    • The best matching chunk score is used for each document
    • Full documents (not chunks) are returned with their complete content
    • The best matching chunk text is shown as a preview highlight

Why Chunking?

Chunking improves search accuracy because:

Example Flow:

User Query: "How to troubleshoot charging errors"
       ↓
Query Embedding: [0.123, 0.456, 0.789, ...] (1536 numbers)
       ↓
Search Chunks Index (3000+ chunks from all documents)
       ↓
Best Matching Chunks:
  - Chunk #5 from "Charging Station Guide" (score: 0.95)
  - Chunk #12 from "Error Code Reference" (score: 0.87)
  - Chunk #3 from "Charging Station Guide" (score: 0.82)
       ↓
Group by Document:
  - "Charging Station Guide": best score 0.95
  - "Error Code Reference": best score 0.87
       ↓
Fetch Full Documents (entire content, not just chunks)
       ↓
Return: Complete "Charging Station Guide" + "Error Code Reference" documents

This approach combines the precision of chunk-level semantic search with the utility of returning full documents for complete context.

Source Filtering

Official Sources Filter

When you need only official Monta content, use the official=true parameter to filter results to company-authored sources.

What Are Official Sources?

The official filter restricts results to content created by Monta:

When to Use Official Filtering

Use official=true when:

Don't use official filtering when:

Example Usage

MCP Tool:

{
  "query": "How does load balancing work?",
  "official": true,
  "maxResults": 10
}

REST API:

# GET request
curl "https://knowledgebase.monta.app/api/query?query=load%20balancing&official=true"

# POST request
curl -X POST https://knowledgebase.monta.app/api/query \
  -H "Content-Type: application/json" \
  -d '{
    "query": "load balancing",
    "official": true,
    "maxResults": 10
  }'

What Gets Filtered Out

With official=true, these sources are excluded:

This ensures you get only Monta-authored content.

Note: Gong sales calls ARE included in official sources because they contain curated pre-sales and post-sales Q&A, unlike Aircall which contains user-generated support call transcripts.

Factual Sources Filter

When you need reliable, documented information (including industry standards), use the factual=true parameter to filter results to trusted sources.

What Are Factual Sources?

The factual filter restricts results to:

When to Use Factual Filtering

Use factual=true when:

Don't use factual filtering when:

Example Usage

MCP Tool:

{
  "query": "OCPP 2.0.1 specifications",
  "factual": true,
  "maxResults": 10
}

REST API:

# GET request
curl "https://knowledgebase.monta.app/api/query?query=OCPP%20specifications&factual=true"

# POST request
curl -X POST https://knowledgebase.monta.app/api/query \
  -H "Content-Type: application/json" \
  -d '{
    "query": "OCPP specifications",
    "factual": true,
    "maxResults": 10
  }'

What Gets Filtered Out

With factual=true, these sources are excluded:

This ensures you get documented, reliable information from both Monta and industry sources.

Note: Gong sales calls ARE included in factual sources because they contain curated pre-sales and post-sales Q&A with confidence scoring.

How Relevance Ranking Works

Results are ranked using multiple factors:

1. Where the Match Appears

Matches in different locations have different importance:

Example: Searching for "solar panels" will rank a document titled "Solar Panel Installation Guide" higher than one that mentions solar panels only in the content.

2. Document Source Quality

Different content sources have different weights:

3. Content Freshness

Newer content ranks higher for time-sensitive sources:

Older content remains available but ranks lower when similar newer content exists.

4. Content Quality

Documents are scored based on metadata:

Via Interactive Tester

Visit /query for an interactive search interface where you can: