Knowledgebase Guide

Building Your RFI Knowledgebase

The quality of every AI-drafted response depends entirely on what is in your knowledgebase. Here is what to include, how to structure it, and how retrieval actually works.

Retrieval

How Retrieval Works

When an RFI arrives, the workflow does not send the full KB to the AI model. That would be slow, expensive, and unfocused. Instead, it runs a semantic search to retrieve the most relevant passages — and only those passages go into the draft context.

The Retrieval Pipeline

  1. Query construction: The workflow builds a search query from the RFI text plus the classification metadata (trade, spec section, question type).
  2. Embedding: Both the query and every passage in the KB are converted to vector embeddings — numerical representations of meaning.
  3. Similarity search: The system finds the KB passages with embeddings closest to the query embedding. Closeness = semantic relevance, not keyword overlap.
  4. Relevance threshold: Passages below the minimum relevance score (default 0.65) are excluded, even if they are the "closest" matches.
  5. Context window assembly: The top-N passages (default 5) are assembled into the draft context with their source citations.
  6. Draft generation: The AI model writes the response using only the retrieved passages as its factual source. It is explicitly instructed not to use general knowledge.
What to Include

What to Include: Content Types

Five types of content consistently produce the highest-quality RFI drafts. Index all five before your first project goes live.

📄

Project Specifications (Project Manual / Spec Books)

The most important content type. Spec books contain the authoritative technical requirements for every trade and system. They are what reviewers consult when answering RFIs, so they are what the AI should consult too.

What to include

  • Division 00–01 (General Requirements)
  • Technical specifications for all relevant trades
  • Referenced standards (ASTM, ASME, ANSI, etc.)
  • Addenda and specification bulletins
💡
Tip: Upload each division as a separate document. Smaller, focused documents retrieve more precisely than one 800-page combined PDF.
📚

Prior RFI Archive

Answered RFIs from this and similar past projects are your highest-signal content. When a new RFI resembles a previously answered one, the AI can draft a response grounded in the prior answer — and the reviewer can reference the precedent.

What to include

  • Closed RFIs with approved responses
  • Both the question and the answer (as a pair)
  • Source project and date for context
  • Any attachments that were part of the response
Don't include: Rejected drafts or informal email discussions that never became official responses. Noise in the archive degrades retrieval quality.

FAQ Documents

Pre-written answers to questions that come up repeatedly — from owners, subcontractors, or reviewers during project kick-off. These are the fastest path to consistent responses for common questions.

What to include

  • Owner-provided FAQ documents
  • Pre-bid Q&A from the procurement phase
  • Design intent clarification memos
  • Pre-construction meeting minutes with decisions recorded
💡
Format matters: Explicit Q&A format ("Q: ... A: ...") retrieves better than narrative memos with the same information buried in paragraphs.
📜

Referenced Standards Excerpts

Specs reference ASME, ASTM, ANSI, NFPA, and other standards by number. When an RFI asks about compliance with a referenced standard, the AI cannot answer well unless the relevant sections of that standard are in the KB.

What to include

  • Sections of referenced standards that are directly cited in the specs
  • Tables and figures from standards that affect design decisions
  • Interpretation bulletins from standards bodies
Do not upload entire standards. They are large, most content is irrelevant, and it degrades retrieval precision. Extract and upload only the sections cited in your project specs.
📝

Design Intent Documentation

Architect and engineer narratives that explain why design decisions were made — not just what they are. When a conflict arises, understanding the design intent helps the reviewer find the resolution that preserves the project's goals.

What to include

  • Basis of Design narratives
  • Architect's project description
  • System narrative descriptions from engineers
  • Value engineering decision records
💡
Often overlooked: Design intent documents are rarely uploaded to traditional RFI management tools. They are high-value content for resolving "why" questions that specs alone cannot answer.
What to Exclude

What to Leave Out

Content That Degrades Retrieval Quality

  • Draft or superseded documents — Old spec versions or preliminary drawings create conflicting results. Always version-stamp documents and remove superseded versions.
  • Informal communications — Email threads, chat logs, and meeting notes without formal decisions recorded. These introduce ambiguity, not clarity.
  • Non-project-specific boilerplate — Generic spec sections without project customization. The AI cannot distinguish boilerplate from project-specific requirements.
  • Scanned documents without OCR — Image-only PDFs cannot be indexed. Run OCR before upload or the document will silently produce no retrievable passages.
  • Entire reference standards — As noted above: extract relevant sections only.
Structure

How to Structure Your KB for Best Retrieval

Tag by Trade and Subject

Every document should be tagged with the trade(s) it covers (Structural, Mechanical, Electrical, Civil, etc.) and any applicable spec divisions. The classification block uses these tags to focus the search.

One Topic Per Document

A 200-page project manual retrieves less precisely than 12 division-specific documents of 15-20 pages each. Break large documents at logical boundaries before upload.

Use Plain Language Titles

Document titles and descriptions are used in citation display. "DIV 15 HVAC SPECS REV4.pdf" is less useful to a reviewer than "HVAC Mechanical Specifications — Division 15, Revision 4, Issued 2026-03-01".

Version Control Matters

When a spec or drawing is revised, upload the new version and mark the old one as superseded (do not delete it — the audit trail may reference it). The workflow will prefer non-superseded documents in retrieval.

Mark Approved Substitutions

If a substitution was approved during the project (via a formal submittal), add a brief document recording the substitution. Future RFIs about the substituted product will retrieve the approval record.

Feedback Loop From Reviewers

When a reviewer edits an AI draft significantly, the workflow prompts them to add the better answer to the KB. Over time, the KB improves with each project's RFI volume.

Why Semantic

Semantic Search vs. Keyword Matching

Traditional document search systems (and most legacy RFI tools) use keyword matching: they find documents that contain the same words as your query. Semantic search finds documents that mean the same thing, even when different words are used.

Keyword Matching

Query: "pipe support spacing"
Finds: documents containing the exact phrase "pipe support spacing"
Misses: "hanger interval", "support rod spacing per B31.3", "maximum unsupported length for 2-inch NPS"

Semantic Search

Query: "pipe support spacing"
Finds: all of the above, plus documents about "pipe support frequency", "maximum span between hangers", and the relevant ASME B31.3 table — because they all express the same concept.

This matters because RFI questions are written by field personnel and subcontractors using their own vocabulary. The spec was written by engineers using theirs. Keyword matching breaks across that vocabulary gap. Semantic search crosses it.

Worked Example

A Concrete Retrieval Example

Here is an actual retrieval trace for a mechanical RFI, showing how the query, retrieved passages, and draft connect.

Incoming RFI

"On grid line B-7, the duct serving AHU-3 conflicts with a structural beam shown on S-204. The beam is 18 inches deep. The duct is 24x16. We cannot find clearance in the ceiling plenum. Please clarify whether the duct can be rerouted or whether the beam can be relocated."

Classification Output

Trade: Mechanical / Structural  |  Type: Conflict  |  Urgency: High  |  Spec: Division 15 + Structural

Retrieved Passages (top 3 shown)

Score: 0.91  |  Source: Division 15 Mechanical Spec, Section 15810, Para 2.3

"Ductwork shall maintain a minimum clearance of 3 inches from all structural members. Where conflicts occur, the mechanical contractor shall submit an RFI for resolution prior to fabrication. Do not relocate structural members without written authorization from the structural engineer of record."

Score: 0.84  |  Source: Prior RFI Archive, RFI-0087, Grid C-12 (similar conflict, 2025-08)

"RFI-0087 Response: The duct serving AHU-1 at grid C-12 was rerouted per the attached sketch SK-M-014. The structural beam at C-12 is a primary lateral frame member and cannot be relocated. Coordinate with structural for revised duct routing."

Score: 0.79  |  Source: Structural General Notes, Sheet S-001

"Primary lateral frame members (designated LF on plans) may not be relocated or modified without written approval from the structural engineer of record. Secondary beams (designated SB on plans) may be relocated subject to a structural review and redesign."

AI Draft Excerpt

"Per Division 15 Section 15810 Para 2.3, ductwork must maintain a minimum 3-inch clearance from structural members and structural members may not be relocated without written authorization from the structural engineer of record [Source 1]. The beam at B-7 should be confirmed as either a primary lateral frame member (LF designation) or a secondary beam (SB designation) per Sheet S-001 general notes [Source 3]. If LF, the duct must be rerouted. A similar conflict at grid C-12 was resolved by rerouting the duct per sketch SK-M-014 [Source 2, RFI-0087]. We recommend confirming beam designation and submitting a proposed rerouting sketch for review."
💡
The reviewer's job is now to confirm the beam designation (takes 30 seconds to check S-204) and either approve the draft as written or update it with the confirmed designation. Compare this to starting from a blank page.
Quick Start

Build Your KB in 30 Minutes

1

Gather Your Documents

Project spec book (by division), any approved substitution records, and the most recent completed RFI log from a similar past project. That is enough to start.

2

Upload and Tag

Upload each document to the Knowledgebase section of your RenderDraw project. Add trade tags and a plain-language description. The system handles OCR, chunking, and embedding automatically.

3

Test With a Sample RFI

Submit a test RFI using a real question from a past project where you know the answer. Check whether the retrieved passages match what you would have searched for manually. If not, add the missing document.

Ready to build your knowledgebase?

Start your free trial and index your first project documents in under 30 minutes.