Radiant Knowledge Library

How your organization calculates key metrics, and which data fields to use. This prevents the AI from picking a plausible-looking field that actually gives wrong numbers.

Examples:

• "For pipeline analysis, use the deal amount field — our ARR field is only populated on closed and renewal deals, not open pipeline."
• "Revenue has multiple components: platform fees, professional services, and add-on products. Don't sum them unless asked for total contract value."
• "When someone asks about 'deal size,' they mean total deal value. When they ask about 'ARR,' they mean the recurring portion on closed deals only."

Fiscal year boundaries, quarter definitions, and labeling conventions. Without this, the AI defaults to calendar year — and every time-based answer is wrong.

Examples:

• "Our fiscal year starts in February. FY2027 runs Feb 2026 – Jan 2027."
• "Fiscal year labels use the end-year, not the start-year. So 'FY2027 Q1' means Feb–Apr 2026."
• "Quarter boundaries: Q1 = Feb–Apr, Q2 = May–Jul, Q3 = Aug–Oct, Q4 = Nov–Jan."
• "When someone says 'this quarter' they mean the current fiscal quarter, not the calendar quarter."

How entities are categorized in your CRM, especially where labels are misleading. This is critical because CRM fields often don't mean what they appear to mean.

Examples:

• "The 'Customer' type field doesn't capture all real customers. Filter by accounts that have a provisioned org ID for an accurate count."
• "Stage names like 'Discovery' appear in both New Business and Expansion pipelines. Always filter by record type when analyzing stages."
• "We have two terminal-loss stages: 'Closed Lost' and 'Rejected.' They use different reason fields — don't mix them up."
• "Account type is unreliable for segmentation — nearly half of accounts have no type assigned."

Which data comes from which system, and what each system covers. Many organizations use multiple CRMs and tools, and each has only part of the picture.

Examples:

• "Meeting data spans three sources: CRM events, call recording tools, and marketing automation. Querying just one gives an incomplete count."
• "All email marketing engagement comes from our marketing automation platform. Event campaigns come from the CRM. Don't look in one system for the other's data."
• "Person records are unified from multiple sources — marketing contacts, CRM leads, CRM contacts, and call participants. Filtering by one source silently excludes people known only through the others."

Internal shorthand, abbreviations, and team-specific language that the AI wouldn't know without being told.

Examples:

• "'PS' in opportunity titles stands for 'Professional Services' — it's a product line, not a generic abbreviation."
• "When the team says 'pipeline,' they mean open New Business opportunities. Renewals are tracked separately."
• "'Partner program' refers to our co-sell motion with channel partners — not a technical integration partnership."
• "'The webinar' without further context refers to our flagship quarterly industry event, not a one-off session."

Known gaps, quirks, or traps in your data. These are the facts that save the AI from producing answers that look right but are subtly wrong.

Examples:

• "Lead source is NULL on the vast majority of opportunities — don't use it for attribution. Use the touchpoint data instead."
• "Account source mostly reflects data-enrichment vendors, not marketing channels. It doesn't answer 'where did this account come from.'"
• "Event engagement is split across multiple sub-channels (conferences, dinners, meetups, etc.). Querying a single event type won't capture the full picture."
• "All our revenue data is in USD. Don't add currency conversion logic."