Spreadsheets and databases hold some of the most precise, high-value knowledge in your organization — but generic AI can't read a table without losing the relational structure that makes it meaningful. RenderDraw's Structured Data KB preserves row-column relationships, supports exact lookups and range queries, and stays live-synced with your ERP and CRM so every workflow always uses current data.
When a Document KB ingests an Excel file, it treats the file as prose and extracts the text row by row, losing the structural relationships between columns. The result is chunks that read like: "Part Number SKU-4892 Diameter 2 inches Material Carbon Steel Price 14.50 Quantity Break 100 Price 12.25 Quantity Break 500 Price 10.80." Semantically, this is garbage — the association between "Carbon Steel" and "2 inches" and "$10.80 at 500 units" is lost.
Structured Data KBs solve this by preserving the row-as-record model. Each row is ingested as a complete, coherent unit with named attributes. The AI block can query: "What is the price of a 2-inch carbon steel part at a quantity of 500?" and receive back a structured record: {"part_number": "SKU-4892", "diameter": "2in", "material": "Carbon Steel", "qty_500_price": 10.80}. The answer is exact, the units are correct, and the attribute relationships are intact.
This distinction has enormous practical consequences for workflows that generate quotes, takeoff cost estimates, or purchase order verifications. A workflow using a properly configured Structured Data KB will produce accurate line-item pricing. A workflow using a Document KB with a flattened Excel file will frequently produce incorrect numbers — with high confidence, because the retrieval "found" something that looked relevant.
The most common source of AI pricing errors in construction workflows is not a bad AI model — it is a pricing table ingested through the wrong KB type. Use Structured Data for anything with rows and columns.
Upload a file, connect a live database, or pull from your ERP. All three stay current through automatic sync.
File uploads are versioned. Each new upload of the same table supersedes the previous version; the old version is archived, not deleted.
Live connectors require a read-only service account. RenderDraw never writes to connected databases.
Raw spreadsheet column headers are often ambiguous, abbreviated, or inconsistent across files from different sources or time periods. Column mapping lets you create semantic aliases for each column — translating your actual headers into meaning that the retrieval system and the AI model can use precisely.
For example, a pricing file might have a column named "QTY_BRK_3_PRC" in the raw export. You map this to a semantic alias: unit_price_500 with label "Unit price at quantity 500 or above" and data type "currency (USD)." The retrieval system now knows exactly what this column means and can answer "what is the price at quantity 500?" correctly by matching "quantity 500" to the unit_price_500 column definition.
Schema definition goes further: you specify data types (string, integer, decimal, boolean, date, currency), units (inches, millimeters, pounds, kg, USD, EUR), and whether each column is a primary key, a foreign key reference, or an attribute. Primary key columns are used for exact-match lookups. Foreign key relationships link tables across multiple sheets or source files — enabling a single query to span a pricing table and a product catalog simultaneously.
Schema definitions are reusable. Define a schema once for your standard pricing table format, and every new pricing file upload that matches that format is automatically mapped — no manual re-configuration per upload.
{
"raw_header": "QTY_BRK_3_PRC",
"semantic_alias": "unit_price_500",
"label": "Unit price at qty 500+",
"data_type": "currency",
"currency": "USD",
"nullable": false
},
{
"raw_header": "MAT_CD",
"semantic_alias": "material",
"label": "Material type",
"data_type": "string",
"enum_values": [
"Carbon Steel",
"Stainless 304",
"Stainless 316",
"Brass",
"PVC"
]
}Structured Data KBs support two retrieval modes, which can be used independently or in combination depending on the workflow task.
Semantic mode treats each row as a small document — its column values are serialized into a structured text representation, embedded, and indexed as a vector. Queries are natural language: "find the lowest-cost stainless steel pipe fitting under two inches" or "what materials are available for flanges in the 150-pound rating class?"
Semantic mode excels at discovery and exploration tasks: "what options do we have for corrosion-resistant fittings?" or "show me all products that match this spec's requirements." It handles synonym resolution, unit conversion (the query asks for "2 inches," the data has "50mm" — the embedding resolves the equivalence), and incomplete attribute matching.
Semantic mode returns a ranked list of rows ordered by relevance. Confidence thresholds apply exactly as they do for Document KBs. Use topK=3–5 for semantic queries in production workflows.
Exact mode bypasses the vector index entirely and runs a deterministic lookup against the structured row store. The workflow AI block (or the workflow config) specifies the lookup criteria as a structured filter: {"material": "Carbon Steel", "diameter": "2in", "rating": "150lb"}.
Exact mode returns only rows that precisely match all specified criteria. This is the right mode when the query parameters are known exactly — for example, in a takeoff workflow that has already extracted the material spec and diameter from a drawing and needs the exact price for that item. No ambiguity, no ranking, no confidence threshold.
Range filters extend exact mode: {"quantity": {"gte": 500}, "material": "Carbon Steel"} returns all rows matching the material with quantity breaks at or above 500. Use range filters for quantity-break pricing lookups, lead-time queries ("items with delivery < 5 days"), and budget ceiling checks.
Hybrid mode combines semantic and exact matching in a single query. The exact filter narrows the candidate set — restricting to a specific product category or material type — and the semantic ranking then orders the filtered results by relevance to the natural language query. This is the most powerful mode for complex lookups: "from our carbon steel fittings catalog, find the most cost-effective option for high-pressure steam applications up to 600°F."
Hybrid mode is the default when a workflow AI block receives a mix of structured parameters (extracted from the source document by a prior workflow step) and free-text query intent. RenderDraw automatically detects when both signal types are available and routes to hybrid mode.
The most valuable structured data — your real pricing, your current inventory, your live product catalog — lives in your ERP or CRM. Static file exports are always stale the moment they're generated. RenderDraw's live sync connectors pull directly from source systems on a configurable schedule, ensuring every workflow uses current data.
Live sync connections are configured with a read-only service account or API credential. RenderDraw supports both polling-based sync (query the source on schedule) and push-based sync (source system calls RenderDraw's webhook when records change). Push-based sync, where available (Salesforce, most modern CRMs), produces near-real-time KB currency with minimal query load on the source system.
Schema mapping is applied at sync time — the same column definitions you configure for a static upload apply to live sync results. If the source system adds or renames columns, RenderDraw flags the schema drift in the KB management panel and pauses sync until the mapping is updated. This prevents silently incorrect data from entering the KB.
Every sync run produces a change report: records added, records updated, records deleted (flagged, not removed from KB). The change report is available in the KB management panel and via the GET /api/knowledgebases/{kb_id}/sync-history API. All workflow outputs generated in a given time window can be cross-referenced against the KB state at the time of generation — essential for quote audits and pricing dispute resolution.
Connect SAP pricing conditions (tables KONP and KOND) via the SAP OData API. RenderDraw maps condition records to a normalized pricing schema:
{
"source": "sap_odata",
"endpoint": "/sap/opu/odata/sap/SD_SLS_OA",
"entities": ["PricingConditionSet"],
"sync_schedule": "daily_02:00_utc",
"schema_map": {
"ConditionType": "pricing_type",
"ConditionScaleQty": "quantity_break",
"ConditionRateValue": "unit_price",
"Currency": "currency",
"ValidityStartDate": "valid_from"
}
}How structured data knowledgebases power real workflow scenarios across construction, manufacturing, and industrial sales.
A takeoff workflow extracts material quantities from a drawing. Each line item fires an exact-match query against the Pricing KB: {"item_code": "SS316-EL-2IN", "qty": 48}. The KB returns the correct quantity-break price. The workflow assembles a line-item cost estimate without any manual pricing lookup.
An RFP response workflow uses a semantic query against the Product Catalog KB to find the best-fit product for the specified application, then an exact query against the Pricing KB for the matched product's unit cost and lead time. Both results are injected into the AI context to generate a precisely priced technical proposal section.
A qualification workflow queries the Approved Vendor List KB (synced daily from the ERP vendor master) to verify that a named subcontractor appears on the approved list for the specified trade category. The query result gates the workflow — unapproved vendors trigger a human-review step before the workflow continues.
A field service workflow queries the Spare Parts KB with the equipment model number and symptom description. The KB returns matching replacement part records — part number, availability, unit cost, lead time — which are presented to the field technician in a structured parts list without requiring ERP system access in the field.
Create your first Structured Data KB, configure the schema, and connect it to a workflow block in 30 minutes.
Read →How structured data retrieval results are injected into the AI context alongside document chunks at runtime.
Read →Compare Structured Data KBs to the four other knowledgebase types and understand when to use each one.
Read →