Creating Knowledge

MergeBase transforms your tables, documents, and connected data into an intelligent, searchable layer that powers AI agents and workflows. This guide walks you through creating and organizing knowledge effectively.

What Counts as Knowledge?

Knowledge in MergeBase lives in tables. That includes:

• Native tables you define yourself, with a custom schema
• Reference tables (thresholds, pricing, validation rules)
• Uploaded files parsed into structured records
• Connected data from external systems (CRMs, databases, SaaS apps)
• Human-in-the-Loop task data, exposed as its own virtual table

Once created, this content can be queried by agents, referenced by workflows, or enriched with AI-proposed columns, without needing to manually search.

Two Ways to Create a Table

From the Create Table button, you can:

1. Upload a File

Use this for static content or a one-time import.

Best for:

• PDF, Word, or text documents
• CSV or Excel spreadsheets
• Audio files
• Policy documents, templates, archived reports

Steps:

1. Click Create Table > Upload file
2. Choose the file(s) via drag-and-drop or the file browser
3. Monitor parsing progress with real-time indicators
4. The parsed content becomes a new table in MergeBase

Supported file types: PDF, CSV, Excel, Word, and audio.

2. Connect a Source

Use this for dynamic data that should stay in sync with an external system — whether it's a database or a SaaS app. MergeBase uses a single connect flow for both, backed by 30+ connectors.

Best for:

• CRM records (Salesforce, HubSpot)
• Databases (Postgres, Snowflake)
• Productivity & collaboration apps (Google Sheets, Google Drive, Notion, Airtable, SharePoint, Outlook, Teams, Gmail)

Steps:

1. Click Create Table > Connect source
2. Select the system you want to connect
3. Authenticate (API credentials or OAuth, depending on the connector)
4. Choose which tables, sheets, or objects to sync
5. Choose a sync mode — Incremental (only new/changed records) or Full (rebuild everything)

Sync behavior: Scheduled or on-demand sync with status monitoring (Succeeded, Running, Pending, Failed) and error detail on failure.

Configuring a Table for AI Search

Every table — however it was created — needs its fields configured before it's useful for AI retrieval.

Why Configure Fields?

• Control exactly which fields are embedded for semantic search
• Keep structured fields (status, category, date) usable for filtering without polluting search relevance
• Avoid wasted processing on fields that add no retrieval value (IDs, timestamps, raw numeric values)

How to Configure:

1. Open the table and click Configure AI Search
2. For each field, assign a role:
   • Vectorize — embedded for semantic (AI) search
   • Filter — used for exact-match structured filtering, not embedded
   • Display — shown in results but not searched or filtered
   • Ignore — excluded entirely
3. Save the configuration to trigger processing

Advanced: Structured Tables for Q&A

Use this approach when your knowledge comes in a question-answer or rule-based format:

Recommended structure:

• Question column: Common queries or scenarios (Vectorize)
• Answer column: Detailed responses or guidance (Vectorize)
• Category/Tag column: For filtering (Filter)
• Priority column: For handling multiple applicable answers (Filter or Display)
• Last Updated column: For content freshness tracking (Display)

Tips for effectiveness:

• Use consistent phrasing and terminology
• Include variations of common questions
• Keep answers concise but complete

Enriching Existing Tables with AI

Rather than manually adding derived columns, you can describe what you want in plain language using AI Column: type a prompt (e.g. "sum total amounts for matching GL Codes"), review the AI-proposed formula and a live preview of the calculated values, then confirm to add it as a real column. See the Enrichment Tool doc for the full workflow.

Quality Management

Keep your tables reliable and current:

Regular maintenance:

• Periodic reviews: Audit content for accuracy and relevance
• Sync health: Watch for tables stuck in a Failed sync state and fix the underlying connector/auth issue
• Field role review: Revisit Vectorize/Filter/Display/Ignore assignments as your table's schema evolves
• Owner validation: Have team owners review and approve new tables

Content quality checklist:

• Is the information current and accurate?
• Is the table's schema and naming clear and searchable?
• Are fields assigned the right role for how they're actually used?
• Does the content support good chunking (clear structure, not one giant blob per row)?
• Are there any duplicate or conflicting tables?

Best Practices Summary

For maximum effectiveness:

• Start with high-impact, frequently-used tables
• Use consistent naming conventions across all tables
• Only mark long-form, descriptive fields as Vectorize — use Filter for exact-match fields
• Regularly validate content accuracy with subject matter experts
• Test retrieval in realistic scenarios before relying on it in production workflows