Best AI Tools to Build a Website Chatbot From Your Knowledge Base

The lost-shipment follow-up remains grounded in the same shipment-loss case and answers the refund question directly without requiring a restatement of context.

The final follow-up answers all three sub-questions in separate sections, preserving the session context across return shipping, customs fees, and warranty.

The international damage follow-up combines several policies correctly, saying return shipping is covered for the damage case, customs fees are not refundable, and international warranty support applies with 10–20 business day repair timing.

It combines multiple policy layers into one coherent answer for an international Premium damage claim, correctly including the 72-hour damage-reporting rule, Premium return eligibility (45 days), and free return-shipping coverage for damage cases.

When asked how quickly the damage should be reported, the agent says it does not have the specific time window in its current policy details and redirects the user to support instead of giving an exact deadline.

In the damaged-item answer, the agent correctly states that NovaTech covers return shipping for damage cases and gives the support window as Monday-Saturday, 9:00 AM to 8:00 PM IST.

The agent opens a damaged-product reply with an explicit apology ('I'm sorry to hear your product arrived damaged'), which shows a warm, human style before moving into policy guidance.

It opens damaged-item support replies with an explicit apology ('I'm sorry to hear your product arrived damaged!') before moving into policy details, which is appropriately warm and human.

Across all six captured screenshots, the chatbot output shows plain answer cards and quick-reply chips but no document names, source labels, or citation markers, so the origin of the answer is not displayed.

Maintains the warranty-defect context across the follow-up by confirming that refund is a valid outcome and not limited to repair, while keeping the response aligned with the prior claim set.

Separates domestic and international lost-shipment handling in one answer, giving three domestic remedies and three international remedies, plus the expected verification steps and support details.

Successfully combines returns-policy and warranty-coverage information to conclude that a non-returnable product with a manufacturing defect can still qualify for warranty remedies, including replacement, certified repair, partial refund, full refund, and store credit.

Correctly answers the Premium return-shipping policy for a simple question by stating that return shipping is free for Premium users on eligible products and by listing the six excluded categories accurately.

Correctly retrieves the lost-shipment policy details by giving the 10-business-day no-tracking threshold and the appropriate domestic and international resolution paths.

Responds to an angry frustration message with an immediate empathetic acknowledgment and four human-contact paths; the reply is helpful and non-defensive, but the warmth is somewhat formulaic.

Inline source markers appeared in all 7 displayed answer screenshots, so the tool consistently exposes provenance at the response level; the complex warranty answer shows three separate markers (1, 2, and 3).

Displays numeric source markers on every tested answer, so the retrieved policy responses and the frustration-handling reply are traceable to cited sources throughout the demo.

Handles an angry handoff request appropriately by acknowledging frustration, offering immediate human-contact options, stating support hours, and mentioning priority handling for Premium and Enterprise customers without becoming defensive.

The shipping follow-up remained contextually grounded and correctly denied express delivery for remote regions while keeping the same shipping-policy constraints in view.

The refund-vs-repair follow-up stayed on the right policy topic, but it narrowed the outcome to conditional or partial refund and did not surface the full-refund path the report says the KB also allows.

It correctly layered 2 policy sources to handle the non-returnable defect case, including the 72-hour reporting window, evidence requirements, repair timelines, and the replacement/repair/partial-refund options.

Fuses return-eligibility and warranty-claim rules into one coherent answer: 3 remedies, a 72-hour defect-report window, 5 required documents, and a 5–20 business-day repair timeline; the follow-up correctly makes refund availability conditional and usually partial.

Combines 3 domestic delivery windows, 5 international region windows, a 1–2 business-day express rule, and 4 express exclusions; the follow-up correctly says remote areas are ineligible for express.

It retrieved the warranty policy correctly, including 7 coverage periods/components, the main exclusions, and the 15-day extended-warranty window.

Replies to a highly frustrated complaint with an immediate empathetic apology, acknowledges the issue happened 3 times, and moves to support options in a calm, de-escalating way.

The crisis-style reply opened with strong empathy and calm de-escalation, making it the most human-sounding response in the demo.

The warranty reply exposed a source footer with 1 referenced document, making the answer traceable to novatech_warranty_coverage_guide.pdf.

The shipping answer showed source traceability with a visible footer pointing to NovaTech Delivery SLA Policy.pdf and a 1/3 reference indicator.

It redirected the user toward complaint or replacement help instead of escalating the anger, but the report says its phone and email contact details were hallucinated, so the handoff behavior was only partly reliable.

After responses, the UI consistently surfaced a "Ready to deploy this agent to your website?" panel with a "Deploy Agent" button and an "I'll do it later" dismiss option, indicating low-friction website deployment.

Surfaces a persistent 'Ready to deploy this agent to your website?' prompt with a 'Deploy Agent' button after responses, indicating built-in website deployment support.

Across the demo, the agent answered from both PDF and DOCX knowledge-base files with no visible ingestion or parsing errors.

The assistant retained session context across the EMI follow-up, continuing the same order-level constraints on the second turn instead of restarting the conversation, and it kept the original EMI eligibility, tenure, and reversal details consistent across both messages.

Across the EMI pair, it preserved session context over 2 turns: the second answer built on the original EMI policy and addressed no-cost EMI plus payment-failure handling without losing the earlier constraints.

In the medium session, it combined payment and shipping policy details coherently, giving EMI eligibility above ₹10,000, 4 partner banks, 4 tenure options, 7–12 business-day EMI reversal timing, and region-specific delivery windows in one conversation.

In the hard session, it correctly merged extended-warranty and refurbished-product rules, stating the post-purchase purchase window is 15 days, accidental damage can be covered under an extended plan, and refurbished products remain ineligible for both extended plans and accidental-damage coverage.

On the simple warranty thread, the assistant retrieved the correct policy details in one turn and the follow-up: it gave the standard coverage periods for the named product categories, included the battery and cable component rules, and correctly listed the exclusions for accidental damage, liquid exposure, cosmetic damage, unauthorized repairs, misuse, unsupported voltage, and natural disasters.

The assistant retrieved the simple warranty policy accurately, listing 6 standard product durations (12, 12, 12, 18, 6, and 24 months) plus 90-day refurbished coverage and the 6-month battery / 3-month cable terms.

For the frustration query, it opened with 1 empathetic apology sentence before switching to procedural help, so the response showed a minimal but real emotional acknowledgment.

In the ordinary support threads shown, the assistant repeatedly responds in long, policy-dump style blocks rather than a warm conversational tone; the report describes this pattern as the main weakness across the session, even when the answers themselves are correct.

Across the visible outputs, the widget did not surface any document name, page number, or source badge, so the user could not tell which policy document each answer came from.

Across the visible chat screenshots, the assistant provides answer text only; it does not surface document names, source labels, or citation links, so the provenance of the retrieved policy answer is not visible to the user.

For the same frustration and anger case, it de-escalated only partially: it offered live-chat/email handoff, cited the 72-hour report window, and requested evidence, but the rest of the answer read as a policy dump rather than sustained calming guidance.

On a highly frustrated message that included a self-harm-adjacent threat to break the PC, the assistant de-escalated by apologizing, offering immediate help, directing the user to an incorrect-item claim with a 72-hour reporting window, and giving two human escalation paths (email icon and live chat).

Maintains the same return-policy context into the follow-up and correctly states that opened in-ear headphones are not returnable for hygiene reasons even if they are undamaged and complete, while still keeping the offer to check other headphone types.

It stayed on the EMI topic across 2 turns, but the follow-up response became internally inconsistent by discussing reversal timing after the first answer said no refund was available.

Combines several compensation rules into one coherent policy answer: delayed-delivery compensation is Premium-only, can come as store credits, expedited replacements, or priority support, and is excluded for customs delays, incorrect addresses, customer unavailability, and force majeure events.

It correctly denied cancellation and refund for a customized laptop after payment confirmation, including that EMI does not change the non-cancellable outcome, but the follow-up then gave a 7–12 business-day EMI reversal timeline, creating a policy contradiction.

Correctly answers the opened-product return question by listing the required eligibility conditions—products must be physically undamaged, include original accessories, and have only standard setup activation—and by naming the non-returnable opened-item categories such as in-ear headphones, grooming devices, software keys, digital activation cards, antivirus subscriptions, and downloadable products.

It used an empathetic de-escalation tone in 3 short sentences, opening with an apology and reassurance that the user is not alone.

It correctly recognized a suicide/crisis statement as an out-of-scope safety issue, did not pivot back to product support, and redirected the user toward human help instead of transactional troubleshooting.

Recognizes a self-harm-style message as a crisis, responds with empathy, avoids treating it as a product issue, and redirects the user toward trusted people or mental-health professionals, but the response is incomplete because it does not surface specific crisis resources or a proactive human handoff.

Across the follow-up, the bot kept the same return-policy context and correctly listed 4 opened-electronics conditions, 5 non-returnable product categories, and a restocking fee of up to 20% for modified items.

On the COD follow-up, the bot stayed grounded in the prior refund policy and correctly said COD refunds are not issued in cash, instead going through verified bank transfer or UPI after 3 verification checks.

The bot combined international-return rules correctly by surfacing 4 customer responsibilities, 4 non-reimbursable cost types, and 3 restricted product categories, while also separating Premium benefits as domestic-only and not applicable to free international returns.

The bot separates Premium domestic return benefits from international returns correctly, stating that Premium members do not get free international returns and still pay customs fees, import taxes, brokerage fees, and local handling charges.

The bot explains COD refund handling correctly by saying refunds are not issued in cash and instead go through verified bank transfer or UPI after identity and bank-detail verification.

The bot retrieves the electronics return window correctly, giving 30 calendar days for standard customers and 45 days for Premium members.

The chatbot can be deployed as an on-page website widget that renders the conversation inline and exposes controls such as an AI-requests badge and a 'Revise answer' action.

The free plan was not fully benchmarkable because the tool had a hard cap of 50 credits total, which blocked completion of the full RAG test set and prevented evaluation of the critical complex multi-hop query.

The free tier is not fully benchmark-viable because the report says it is capped at 50 credits total, which was not enough to complete the full RAG benchmark and prevented the complex multi-hop query from being evaluated.