Knowledge Base
GPTWeb Collections Strategy — Getting It Right from the Start
the single most impactful thing you can do to improve your GPTWeb AI response quality is to build a thoughtful collections strategy from day one. Collections are not just folders — they are the first stage of GPTWeb's two-phase semantic search engine. When a visitor asks a question, the AI first routes to the right collection before searching individual documents. A well-structured collection strategy means faster, more accurate, more contextually appropriate answers. A poorly structured one means generic, unfocused responses that miss the mark.
The Core Principle — Collections are Semantic Routing Containers
Every collection you create has a name and a description. That description is not just metadata — it is the semantic signal the AI uses to decide which collection to search for any given question. A vague description like 'product stuff' will produce poor routing. A rich, specific description like 'Core product information including features, capabilities, specifications, getting started guides, and how-to tutorials' will route accurately every time. This is the most important best practice in your entire collections strategy: write descriptions as if you are explaining the collection's contents to a new employee on their first day.
Recommended Starter Collection Structure
| Collection Name | Type | Description Best Practice | What to Upload |
|---|---|---|---|
| Product Documentation | Documents | Core product information, features, capabilities, specifications, getting started guides, and how-to tutorials | Feature guides, how-tos, user manuals, onboarding docs |
| Sales & Pricing | Documents | Product pricing, licensing tiers, volume discounts, ROI materials, proposals, and enterprise quotes | Pricing pages, proposal templates, ROI calculators, competitive positioning |
| Support & FAQ | Documents | Customer support articles, troubleshooting guides, frequently asked questions, and known issue resolutions | Help articles, FAQ docs, troubleshooting guides, release notes |
| Product Videos | Videos | Product demos, feature walkthroughs, tutorial videos, and onboarding video content | MP4 demos, tutorial recordings, webinar recordings |
| Product Data | Chart Data | Structured data for pricing tables, feature comparison tables, and performance metrics | CSV or Excel files with pricing data, comparison matrices, benchmark data |
| Product Images | Images | Product screenshots, UI diagrams, feature illustrations, and visual reference assets | PNG/JPG screenshots, architecture diagrams, product photos |
Two-Phase Routing — Why Structure Matters
GPTWeb's Knowledge Base (RAG) engine uses a two-stage hybrid search. Stage 1 compares the visitor's question against your collection descriptions using semantic similarity — only the most relevant collections are selected. Stage 2 searches the documents within those selected collections for the most accurate chunks. This means that if your collections are too broad or too similar in description, Stage 1 routing becomes ambiguous and response quality drops. The goal is clear, distinct collections with non-overlapping descriptions.
Collection Description Quality Guide
| Dimension | Good Example | Bad Example | Why It Matters |
|---|---|---|---|
| Specificity | Technical API documentation, integration guides, SDKs, and code examples for developers | Dev docs | Specific keywords improve semantic matching accuracy in Stage 1 routing |
| Scope | Product pricing, licensing tiers, volume discounts, and enterprise quotes | Pricing stuff | Clear scope prevents content from being retrieved by unrelated queries |
| Content Types | Video walkthroughs, product demos, and onboarding tutorials for new users | Videos | Describing the nature of content helps the AI understand when to route here |
| Audience Signal | Support articles, troubleshooting guides, and FAQ content for existing customers | Help content | Audience context improves routing for intent-specific visitor questions |
| Behavioral Hints | Release notes sorted by date newest to oldest — platform updates and changelog | Updates | Collection descriptions can influence how the AI presents and orders results |
Advanced Strategy — Collection Tuning
RAG Tuning Settings for Collections
| Setting | Default | Lower Value Effect | Higher Value Effect | When to Adjust |
|---|---|---|---|---|
| Collection Threshold | 40% | More collections match — broader context, may include tangential content | Fewer collections match — tighter focus, may miss nearby relevant docs | Lower if answers miss relevant info; raise if answers include irrelevant content |
| Collection Fallback | 25% | Almost always finds some collection — good for broad or vague queries | More likely to return 'I don't have info on that' for off-topic queries | Lower for broad topic sites; raise for tightly scoped product sites |
Best Practices Summary
- Start with 5-6 focused collections rather than 20 broad ones — quality of routing beats quantity of containers
- Write collection descriptions with specific keywords your visitors would actually use in questions
- Keep collection types pure — do not mix videos into a documents collection or images into chart data
- Use AI Auto-Categorize when uploading large content batches — it suggests the right collection based on content analysis
- Add behavioral hints to descriptions where relevant — e.g. 'sorted newest to oldest' for release notes collections
- Review collection routing accuracy after your first 50 visitor conversations and adjust descriptions based on patterns
- Create a dedicated Competitive Intelligence collection if you want your AI to answer questions about the competitive landscape
5-6
Recommended Starter Collections
35-45%
Collection Threshold (Recommended)
10,000+
Max Documents per Tenant
500,000+
Max Chunks per Tenant
A strong collections strategy is the foundation everything else builds on — AI Scoring, AI Campaigns, and DQLs™ all perform better when your RAG engine is routing with precision. Need help planning your collection architecture? Reach out at support@gptweb.com. GPTWeb is the future of engagement, websites, and marketing automation combined — built for the AI era, built for now. 🚀