A RenderDraw knowledgebase is live in under thirty minutes. This guide walks every step — from clicking "New Knowledgebase" to watching your first workflow block retrieve your own company data instead of hallucinating generic answers.
A knowledgebase is only as useful as the content you put into it. Before uploading files, spend five minutes thinking about the retrieval goal: what questions will workflow blocks ask of this KB? What documents, tables, or assets contain the authoritative answers? The clearer the retrieval goal, the better your chunk configuration and context rules will be.
A common mistake is creating one massive, catch-all knowledgebase and expecting the AI to sort it out. In practice, tightly scoped knowledgebases — one for product specs, one for pricing, one for compliance — produce dramatically more precise retrieval than a single monolithic store. You can always combine multiple KBs in a single workflow block using multi-KB queries.
Rule of thumb: If you find yourself saying "this KB has docs about X, Y, and Z" and X, Y, Z are unrelated domains, split it into three KBs. Multi-KB queries handle the cross-domain composition at runtime.
From the platform sidebar, navigate to Workflows → Knowledgebases → New Knowledgebase. You'll be prompted to select the base type: Document, Structured Data, CAD Asset, Response Template, or Compliance. The type you choose sets the default chunking strategy, embedding model, and retrieval mode — all of which you can override in later steps.
For most teams starting out, choose Document for anything that is primarily prose — spec sheets, proposals, technical manuals — and Structured Data for anything tabular like pricing Excel files or BOM CSVs. You can create as many knowledgebases as you need; there is no per-tenant limit.
Once created, RenderDraw assigns the KB a unique identifier (e.g. kb_products_2025q4) that you'll reference when connecting it to workflow blocks. Write this down — it appears in the KB settings panel and in the workflow block editor.
Workflows → Knowledgebases → NewCommon mistake: Selecting "Document" type for an Excel pricing sheet. Use "Structured Data" to preserve row/column relationships instead of flattening them into prose chunks.
Give the knowledgebase a specific, descriptive name. Avoid generic labels like "Documents" or "Data." Names like "Product Spec Library — Conveyor Systems 2025" or "ASHRAE Compliance Standards — HVAC" tell workflow authors exactly what a KB contains when they're building a new automation.
Apply category tags to group related knowledgebases. RenderDraw supports hierarchical tagging: you might tag a KB as domain:products, product-line:conveyors, and year:2025. Tags are used in the workflow block editor to filter available KBs and in the API to query KBs programmatically.
Set the access scope: Tenant-wide makes the KB available to all projects in your organization. Project-scoped restricts it to a single project. Role-scoped limits access by user role. For sensitive information like pricing data or executive proposal templates, project or role scoping is strongly recommended.
RenderDraw supports three ingestion modes. Manual upload lets you drag-and-drop files directly from your desktop — useful for one-time imports of spec sheets, standards documents, or archived proposals. Supported formats include PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), HTML, and plain text.
Live source connection links the KB to an external datasource that syncs on a schedule. Supported sources include SharePoint document libraries, Google Drive folders, and email attachment archives. Once connected, RenderDraw polls for new and updated files and re-ingests them automatically.
API ingestion lets your existing tooling push documents programmatically via the POST /api/knowledgebases/{id}/documents endpoint. This is the preferred mode for teams with automated document pipelines — for example, generating spec PDFs from a CAD system and immediately pushing them into a KB without manual intervention.
All ingested documents are versioned. If you re-upload a file with the same name, the previous version is retained in the version history and remains retrievable for audit purposes. The current version is what workflow blocks retrieve by default.
Uploading scanned PDFs without OCR enabled. Image-only PDFs produce zero retrievable text. Enable OCR in the KB settings or pre-process documents with a text layer before upload.
Chunking is how RenderDraw splits documents into retrievable pieces. The chunking configuration is the most consequential technical decision in knowledgebase setup — it determines the granularity of what the AI can retrieve and how much context surrounds each fact.
Chunk size (measured in tokens) controls how large each retrievable piece is. A chunk size of 512 tokens produces fine-grained retrieval — good for dense technical documents where you want to pull a specific paragraph. A chunk size of 1,024–2,048 tokens produces broader retrieval — good for proposal libraries where a section needs to stay together to make sense.
Chunk overlap is the number of tokens shared between adjacent chunks. Overlap of 10–15% (e.g. 100 tokens overlap on a 512-token chunk) prevents important context from falling into a gap between chunks. For structured data, set overlap to zero — rows do not need to bleed into adjacent rows.
Splitting method determines where chunk boundaries fall. Paragraph splitting respects natural document structure and is the recommended default. Fixed-size splitting divides at exact token counts regardless of sentence boundaries — useful for highly uniform documents. Section splitting uses heading structure (H1/H2/H3) to create section-level chunks — ideal for long technical manuals.
Setting chunk size too large (4,000+ tokens). Large chunks consume most of the AI's context window for a single retrieved piece, leaving no room for multi-KB results and response generation. Keep chunks tight and rely on topK=3–5 to cover breadth.
AI context rules govern how your knowledgebase interacts with workflow AI blocks at runtime. These rules sit between the retrieval layer and the generation layer — they filter, format, and constrain what retrieved content looks like before it reaches the AI model.
Retrieval depth (topK) sets how many chunks are retrieved per query. A topK of 3 returns the three most semantically similar chunks. For narrow, precise documents (a single product datasheet), topK=2 is often sufficient. For broad KB searches (a library of 200 proposals), topK=5–8 ensures sufficient coverage.
Confidence threshold sets the minimum similarity score (0.0–1.0) a chunk must achieve to be included. Setting this to 0.75 means only highly relevant chunks reach the AI. Setting it lower (0.5) broadens retrieval but risks including loosely related content. Start at 0.70 and tune based on output quality.
Citation mode controls whether the AI block inlines source references in its output, appends a footnotes section, or suppresses citations entirely. For compliance-sensitive workflows, always enable citation mode — it is the audit trail that lets a human reviewer verify every claim before approving a response.
Fallback behavior defines what the AI block does when no chunk exceeds the confidence threshold. Options: pass empty context (the AI answers from general knowledge, flagged with a "no KB match" warning), return error (the workflow step fails and triggers a human-review gate), or use secondary KB (fall through to an alternative knowledgebase).
Connecting a KB to a workflow and then never updating the source documents. A knowledgebase without a re-sync schedule is a liability — set one. Even quarterly updates prevent the most damaging stale-data hallucinations (outdated pricing, superseded specs).
With the knowledgebase created, ingested, and configured, the final step is attaching it to one or more workflow blocks. Open the workflow editor, click on an AI block (e.g. "Generate Section Response," "Analyze Technical Requirements," "Lookup Pricing"), and find the Knowledgebase panel in the block properties sidebar.
Select your KB from the dropdown — it will appear by name and tag. For multi-KB queries, click Add Another KB and select additional knowledgebases. You can attach up to five KBs to a single AI block. RenderDraw will fire parallel retrieval queries against all attached KBs and merge the results before injecting them into the AI context.
After attaching, use the Test Retrieval button to run a sample query against your KB without executing the full workflow. This lets you verify that the right chunks are coming back before you run a live document through the automation. The retrieval test panel shows exactly which document, page, and chunk each result came from.
Once satisfied, save the workflow. Every future run of that workflow will automatically query the attached knowledgebases, ground the AI's responses in your content, and include source citations in the output.
When the workflow runs, the AI block constructs a query from the current document context, fires it against all attached KBs simultaneously, ranks and merges results, injects the top chunks into the model context, generates the response, and appends citations — all in under three seconds.
Your knowledgebase is running. Here's where to go deeper.
Understand the five knowledgebase types and when to use each one for optimal retrieval.
Read →The technical deep dive on RAG, semantic search, topK retrieval, and multi-KB composition at runtime.
Read →Supported formats, chunking details, batch upload, API ingestion, and live source sync.
Read →