How to Use Thoughtly Genius for Knowledge-Grounded Conversations

How to Use Thoughtly Genius for Knowledge-Grounded Conversations

Every lead conversion team has the same problem: the AI agent sounds great on scripted questions and falls flat the moment a caller asks something specific—"Do you serve Broward County?" or "Is there a penalty for early payoff?" Genius is Thoughtly’s built-in knowledge base that solves this by giving your voice agents retrieval-augmented access to your actual business information during live calls.

This guide walks through creating a Genius knowledge base, structuring content for reliable retrieval, binding it to your voice agents, and maintaining accuracy over time. By the end, your agents will answer caller questions with company-specific facts instead of generic filler—or worse, hallucinated details.

The examples below use insurance and mortgage scenarios, but Genius applies equally to education enrollment, healthcare scheduling, real estate, home services, automotive, and any other high-volume consumer funnel where callers expect specific answers.

What You’ll Need

• A Thoughtly account on any plan (Genius is available on Flex, Scale, and Enterprise).
• At least one voice agent built in the Agent Builder.
• Your business content ready: FAQs, product/service descriptions, pricing details, coverage areas, eligibility rules, or policy documents. See Genius best practices for content preparation tips.
• Optionally, URLs to public pages you want the agent to reference (pricing pages, service area maps, compliance disclosures).

Step 1: Create a Genius Knowledge Base

Open Tools → Genius in the Thoughtly dashboard and click Create New Genius. Give it a descriptive name that maps to the agent or vertical it will serve—something like "Auto Insurance FAQ" or "Mortgage Product Details" rather than "Knowledge Base 1."

Each Genius database is a self-contained knowledge store. You can connect the same database to multiple agents (for example, your inbound qualification agent and your follow-up re-engagement agent can share one product knowledge base), or maintain separate databases per vertical if the content diverges significantly.

Naming conventions that scale well

• Include the vertical or business unit: "Home Services – Service Area & Pricing."
• Include the content scope: "Medicare Advantage – Plan Comparison Q&A."
• Avoid generic labels. When you have 15 Genius databases six months from now, "KB 3" will not help anyone.

Step 2: Add and Structure Your Data Sources

Select your new Genius database and click Add Data Source. Thoughtly supports four primary source types: Text, PDF, Audio, and URL. A fifth type, CSV, is supported but should be used with caution—structured tabular data has a higher hallucinationHallucinationWhen an LLM-driven agent confidently states something incorrect. Mitigated with RAG, strict prompting, and evals against ground-truth data.Read full definition → risk during RAGRetrieval-Augmented Generation (RAG)A technique that lets the LLM ground its responses in retrieved documents (knowledge base, FAQs) instead of relying on training data alone.Read full definition → retrieval. If you have spreadsheet data, convert it to natural-language Q&A entries first.

Recommended source types by use case

Give each source a descriptive name when you upload it. Processing typically takes 2–10 minutes depending on file size and type. A progress indicator in the Genius source table shows when indexing is complete—wait for it to finish before testing.

Step 3: Format Content for Reliable Retrieval

This is the step most teams rush through, and it is the one that determines whether your agent sounds informed or confused. RAG retrieval works best with natural-language content structured as question-and-answer pairs.

The Q&A format

Write each knowledge entry as a question a real caller would ask, followed by a complete answer your agent should give. This mirrors how the retrieval engine matches caller intent to stored content.

Q: What types of auto insurance do you offer?
A: We offer liability, collision, comprehensive, uninsured motorist,
and personal injury protection (PIP). Liability is required in most
states. We can bundle multiple coverage types for a discount.

Q: How quickly can I get a quote?
A: Most quotes are ready in under five minutes once we have your
vehicle information, driving history, and coverage preferences.

Q: Do you cover drivers under 25?
A: Yes. Drivers under 25 are eligible for all policy types.
Rates are typically higher for drivers under 25 due to actuarial
risk, but we offer good-student and defensive-driving discounts
that can offset the difference.

Content formatting rules

• Keep each answer self-contained. The agent may retrieve a single Q&A pair without surrounding context, so every answer should make sense on its own.
• Use specific, concrete details. "We offer various plans at competitive prices" retrieves poorly. "Our Silver plan is $189/month with a $2,500 deductible" retrieves well.
• Avoid internal jargon unless callers use the same terms. If your team says "PIF" but callers say "paid in full," use the caller’s language in the question and include both in the answer.
• Remove time-sensitive information that will go stale or commit to updating it monthly.
• Do not duplicate the same fact across multiple sources. Conflicting entries increase hallucination risk.

What to avoid

Do not paste raw spreadsheets or CSV tables with complex column relationships. The RAG engine is optimized for natural language, not structured data grids. If you have a rate table with 50 rows, rewrite the most commonly asked combinations as Q&A pairs and keep the full table as a reference document your human team uses offline.

Do not upload entire call recordings without review. Audio sources are transcribed and indexed, which is useful for learning common caller phrasings—but old recordings may contain outdated pricing, discontinued products, or incorrect information that your agent will treat as current fact.

Step 4: Bind Genius to Your Voice Agent

With your knowledge base populated and processed, connect it to your agent. Open the Agent Builder, select the agent you want to augment, and navigate to the Genius tab in the agent’s settings panel.

Select the Genius database you created from the dropdown. Save your changes. The agent immediately has access to the knowledge base—no restart or redeployment required.

Key binding behaviors

• One agent, one Genius. Each agent connects to a single Genius database at a time through the Agent Builder. If you need information from multiple domains, consolidate relevant content into one well-organized database.
• Multiple agents, same Genius. Multiple agents can share one Genius database. This is ideal when your inbound qualification agent and your outbound re-engagement agent need the same product knowledge.
• Hot-swap capable. You can switch which Genius is connected to an agent at any time. The agent immediately starts using the new knowledge base. Test before swapping in production.
• Automation-level override. In Thoughtly Automations, you can select a specific Genius source for individual call steps. This lets you scope the knowledge to a single source, which reduces context noise and hallucination risk for focused workflows.

Remember: Genius provides information, not instructions. It answers caller questions with facts from your knowledge base. Agent behavior, conversation flow, and qualification logic come from your node prompts, outcomes, and agent settings—not from Genius entries.

Step 5: Test Knowledge Retrieval

Before going live, test your Genius-backed agent using Thoughtly’s built-in Test Agent (text chat) and Call Me (real voice call) features.

Testing checklist

• Ask 5–10 questions that should be answered from Genius content. Verify the agent’s responses match your source material.
• Ask 2–3 questions the Genius does not cover. Confirm the agent does not hallucinate an answer—it should handle the gap gracefully (e.g., offering to connect the caller with a human).
• Ask a question using different phrasing than your Q&A entries. If callers say "Am I eligible?" but your entry uses "What are the requirements?", both should retrieve the same content.
• Test edge cases: ambiguous questions, multi-part questions, and questions that span two different sources. These reveal retrieval gaps early.
• Review call transcripts after test calls to see which Genius entries were retrieved and whether the agent cited them accurately.

If the agent consistently misses a question, check whether the content exists in Genius, whether it is formatted as clear Q&A, and whether the phrasing is close enough to caller language. Adding a second Q&A pair with alternate phrasing often fixes retrieval gaps.

Step 6: Maintain and Update Sources Over Time

Genius is not a set-and-forget system. Your business information changes—pricing updates, new service areas, revised eligibility criteria, seasonal promotions—and your knowledge base needs to keep pace.

Maintenance cadence

When you edit a Genius source, Thoughtly re-indexes the updated content and removes the prior indexed version. Wait for processing to complete before testing the agent against updated content. Changes propagate immediately to all connected agents once indexing finishes.

Scaling knowledge across agents

As your agent portfolio grows, resist the temptation to dump everything into one massive Genius database. The sweet spot for optimal retrieval performance is 10–50 well-organized sources. Quality beats quantity—10 precise, well-structured sources outperform 100 hastily uploaded documents.

If you serve multiple verticals (insurance, mortgage, and auto, for example), maintain separate Genius databases per vertical and bind each to the appropriate agents. This keeps retrieval focused and reduces the chance of the agent pulling an insurance answer during a mortgage conversation.

Common Mistakes

• Treating Genius as a prompt. Genius stores factual information for caller questions. It does not control agent behavior, conversation flow, or qualification logic. Use your node prompts and agent settings for instructions; use Genius for answers.
• Uploading raw spreadsheets. CSV and tabular data have higher hallucination rates during RAG retrieval. Convert structured data to natural-language Q&A pairs before uploading.
• Ignoring content freshness. URL sources are fetched once during setup and are not automatically refreshed. If the source page changes, you need to manually re-add or update the source. Set a calendar reminder to review URL-sourced content monthly.
• Overloading the knowledge base. More sources do not mean better answers. Excess, redundant, or loosely related content makes retrieval slower and less accurate. Prune aggressively.
• Duplicating facts across sources. If two sources give slightly different answers to the same question, the agent may return either one unpredictably. Deduplicate ruthlessly.
• Skipping testing. Always test with both expected and unexpected questions after adding or updating sources. Do not assume that uploading correct content guarantees correct retrieval.

Measuring Success

A well-tuned Genius knowledge base should measurably improve your agent’s call quality. Track these metrics:

• Knowledge hit rate. What percentage of caller questions are answered accurately from Genius content? Review call transcripts to identify when the agent retrieved relevant information versus defaulted to a generic response. Target: 85%+ for questions that should be covered.
• Hallucination rate. How often does the agent state something factually incorrect? Compare agent responses against your source material. Any non-zero hallucination rate on critical facts (pricing, eligibility, compliance) needs immediate content remediation.
• Transfer/escalation rate on answerable questions. If callers are being transferred to humans for questions your Genius should cover, that signals a retrieval gap. Track the reason for each transfer and add missing Q&A entries.
• Caller satisfaction and conversion. Agents that give specific, accurate answers convert better than agents that deflect. Compare conversion rates before and after Genius deployment to quantify the impact.
• Content coverage gap rate. Track questions the agent could not answer. If the same question appears across multiple calls, add it to Genius. A shrinking gap rate over time means your knowledge base is maturing.

Frequently Asked Questions

Can I use the same Genius database for both inbound and outbound agents?

Yes. Multiple agents can share one Genius database. This is the recommended approach when agents need the same product or service knowledge. Each agent still maintains its own conversation flow, prompts, and qualification logic independently.

How often should I update my Genius content?

Update immediately whenever business-critical information changes (pricing, eligibility, service areas). Beyond that, review all sources monthly for accuracy and remove anything time-sensitive or outdated. URL sources are fetched once at setup and are not automatically refreshed—add a recurring reminder to re-add or update them.

What happens if the caller asks something not in Genius?

The agent falls back to its base conversation flow and prompts. It will not retrieve irrelevant Genius content just to have something to say. Design your agent’s prompts to handle knowledge gaps gracefully—for example, offering to transfer the caller to a specialist or take a message.

Is there a limit to how much content I can add?

There is no hard published limit, but performance is best with 10–50 well-organized sources. Beyond 100 sources, retrieval accuracy tends to degrade. Focus on quality and relevance over volume.

Can I use Genius to give the agent instructions or change its behavior?

No. Genius is for factual information that the agent retrieves when callers ask questions. Agent behavior, tone, conversation flow, and qualification logic are controlled by your node prompts, outcomes, and agent settings in the Agent Builder. Mixing instructions into Genius content will confuse retrieval and degrade answer quality.

Sources and Further Reading

• Thoughtly Genius — Getting Started — official setup guide for creating and populating Genius knowledge bases.
• Genius Best Practices — content structuring strategies, source type recommendations, and size guidance from Thoughtly docs.
• Genius Agent Binding — connecting, switching, and disconnecting Genius databases from voice agents.
• Genius Data Sources — managing text, PDF, audio, CSV, and URL sources within a Genius database.
• The "Genius" Advantage: How Thoughtly’s AI Learns from Your Business — background on how Genius uses retrieval-augmented generation to ground agent conversations.
• How to Use Thoughtly Variables for Dynamic Call Personalization — related guide on using variables alongside Genius for contextualized agent behavior.
• How to Write Lead-Qualification Prompts Without Overfitting — prompt design guidance that complements Genius knowledge structuring.
• Thoughtly — AI agents that convert the leads you’re already getting by calling, texting, and emailing until someone is ready to talk.