Every vendor in this category will tell you their bot is accurate. Very few will tell you what happens when it does not know something. That gap between the sales pitch and the failure mode is the whole reason this comparison exists.
The comparison covers AI chatbots trained on your website and PDFs, with prices and documentation current to 16 October 2026. The platforms are Intercom Fin, Zendesk AI agents, Tidio Lyro, Chatbase, Botpress, Voiceflow, CustomGPT, RAGFlow, Dify, AnythingLLM and Open WebUI, with ChatPDF, AskYourPDF, Humata, Ada, Forethought, Glean and Salesforce Agentforce covered as a secondary shortlist.
The headline finding is uncomfortable. Of the eleven, one platform gives you a switch that makes the bot refuse to answer rather than guess, and its documentation warns you what happens if you leave it off.
The one property that decides this purchase: will it say it doesn’t know?
Abstention is the act of declining to answer when the answer is not in the source material, rather than generating something plausible from the model’s own training. For a chatbot talking to customers, abstention is the difference between a support agent who says “let me check” and one who invents a refund policy.
RAGFlow is the only platform here with abstention as a named, documented setting. Its quickstart tells administrators to fill in an “Empty response” field: “If you wish to confine RAGFlow’s answers to your datasets, leave a response here. Then when it doesn’t retrieve an answer, it uniformly responds with what you set here. If you wish RAGFlow to improvise when it doesn’t retrieve an answer from your datasets, leave it blank, which may give rise to hallucinations.”
That warning comes from the vendor’s own documentation.
Voiceflow comes second with a “Global no-match” behaviour setting, configurable as either a generated response driven by your instructions or a fixed scripted string. Botpress documents a recipe for the same thing using a Transition Card that fires when the Knowledge Agent returns no answer. Intercom Fin takes a different route: its documentation states that Fin “will hand over if it’s unable to resolve the customer’s question,” and its outcome billing confirms that escalated conversations are not charged.
Everything else relies on you writing the constraint into a prompt. Chatbase’s own best-practices page tells you to add the line “You will only provide answers based on the information from your training data” to the agent instructions. That is advice, not enforcement. An instruction the model can ignore is not a control.
What the documentation establishes
Pricing, features and legal terms reflect vendor pages checked on 16 October 2026. The comparison covers documented citations, fallback behaviour and retrieval controls; it does not include hands-on accuracy or resolution-rate tests.
Which of these eleven can a non-developer actually run?
Six, comfortably. Chatbase’s FAQ says outright that it is “designed for non-technical users” with no coding and copy-paste embed codes. Tidio ships a no-code dashboard and a Playground where you can test Lyro before going live. Intercom Fin and Zendesk are admin-configuration products inside helpdesks you probably already run.
Botpress and Voiceflow use visual builders but expect you to think in tools, triggers and guardrails. Voiceflow’s documentation is explicit that models “respond better to clear, declarative statements than to hedging,” which is a fair expectation of the builder, not of the buyer.
The self-hosted four need competence. What each one asks of the machine:
- Dify: two CPU cores and 4 GB of RAM, the friendliest of the lot.
- AnythingLLM: ships a desktop app and claims “no extra setup required.”
- RAGFlow: four CPU cores, 16 GB of RAM, 50 GB of disk, and Docker 24 or later.
- Open WebUI: an admin surface where you set environment variables for extraction engines and vector databases.
What each platform documents about citations and abstention
This is the comparison that matters most, and it is the one competitor articles skip. Every row below is sourced to that vendor’s own documentation.
| Platform | Customer-facing source links | Documented no-answer behaviour | Where the vendor states it |
|---|---|---|---|
| RAGFlow | Yes, described as “well-founded citations” and “traceable citations” | “Empty response” field returns a fixed string when nothing is retrieved; docs warn blank = hallucinations | Quickstart and README |
| Voiceflow | Yes, “Show source URL(s)” toggle, chat projects only, “Recommended for help centers” | “Global no-match” reply, generative or scripted | Querying the knowledge base; Behaviour |
| Dify | Yes, on by default in chatflows, switchable off in Features | Prompt- and retrieval-threshold-driven; no dedicated abstain switch | Knowledge Retrieval node docs |
| Botpress | Yes, but as a variable: {{turn.KnowledgeAgent.citations}} | “Your bot will generate an appropriate response when it doesn’t find an answer”; Transition Card recipe documented | Knowledge Bases docs |
| Intercom Fin | Partially. PDF and snippet content is private: customers “will not see any link references to them in Fin’s answers” | Hands over to a human; escalated conversations are not billed | Add your content for Fin; Fin AI Agent outcomes |
| Zendesk AI agents | Yes, grounded in help centre and connected external sources | Escalates on “unsupported scenarios”; resolutions are LLM-verified | AI agents page; automated resolutions help article |
| Tidio Lyro | Yes, “in most cases” a link or several, and it can be switched off | Hands off to agents; unanswered questions appear in a Suggestions queue | Data sources help article |
| Chatbase | Not documented as a customer-facing feature | Prompt-based only; docs advise writing the constraint yourself | Data sources; Best practices |
| CustomGPT | Claimed on all plans: “Every response has a link directly to its source” | Not documented | Pricing page (vendor claim) |
| AnythingLLM | Yes, “drag-and-drop uploads and source citations” | Not documented | GitHub README |
| Open WebUI | Yes, “cites where it found the answer” | Not documented | Knowledge Bases docs |
| Ada (secondary shortlist) | Not verified; no public pricing or docs page opened | Vendor claims “built-in safeguards minimize hallucinations” | ada.cx (vendor claim) |
Two of those rows deserve scepticism. CustomGPT’s “Every response has a link directly to its source” and its “#1 for anti-hallucination technology” badge are marketing copy on a pricing page, with no methodology published. Ada’s “built-in safeguards minimize hallucinations” is the same kind of claim. Treat both as unverified.
More importantly, a citation is not a correctness guarantee. A September 2026 arXiv preprint called CiteShade demonstrated that an attacker controlling a single source can induce a model to produce an attacker-chosen wrong answer and attribute it to a trusted source that does not support it.
CiteShade, a preprint submitted to arXiv on 14 September 2026, raised the wrong-answer rate on multi-source multi-hop question answering from 0.01 to 0.68 using one planted source, and found that vulnerability tracks “a model’s propensity to cite rather than its scale.” arXiv:2609.15660. Preprint, not peer reviewed.
The practical reading is uncomfortable for buyers. A model that cites everything confidently is, on that evidence, a model that launders bad answers into well-attributed ones. Where a platform makes you pick which sources are visible to the customer, that control matters.
The counterweight is real too. A June 2026 arXiv preprint on an agentic RAG pipeline deployed at University Medicine Essen had physicians verify 7,326 extracted values against their cited sources; clinicians accepted 96.5% overall, ranging from 80% to 99% by data type. Grounded retrieval with source-level verification does work. The design choices are what separate the two outcomes.
Ingestion: how deep the crawl goes, and what happens to your PDFs
PDF ingestion is where most deployments quietly fail, because the assumption is that a PDF contains text. Three of these eleven will not read a scanned document at all.
Chatbase is the bluntest about it: “The text in a file must be selectable. Scanned PDFs with no text layer cannot be trained on.” That is a clean, documented hard stop. Voiceflow’s data-source table says of .pdf, .txt and .docx files up to 10 MB: “Only text content is imported.” Tidio goes the other way: up to ten PDFs at a time, “no formatting requirements,” and images are understood and used to compose responses even though they are not imported into the Q&A store.
| Platform | Website crawl | Scanned/image-only PDF | Tables in PDFs | Third-party connectors | Re-sync |
|---|---|---|---|---|---|
| Intercom Fin | Public URLs, synced weekly; native Intercom content ingested “almost instantly” | PDF upload supported, text scraped within minutes; no OCR documented | Not documented | Zendesk, Confluence, Guru, Notion, Salesforce, Box, Freshdesk, Document360 | Weekly for web; manual otherwise |
| Zendesk AI agents | Help centre plus “external sources like Google Drive or PDFs” | Not documented | Not documented | Knowledge Graph unifies third-party sources | Not documented |
| Tidio Lyro | “Scan priority pages” capped at 100 pages; up to 500 URLs total, Plus needed beyond | Yes, images understood | Via PDF import | Zendesk Help Center; Intercom and Gorgias tickets via Tidio | Manual, or automatic weekly on Plus and Premium |
| Chatbase | Crawl from homepage, XML sitemap or individual URLs; exclude paths; scraping country | No. Selectable text required | Not documented | Notion; tickets from Salesforce or Zendesk | Auto-resync weekly, only changed pages re-embedded |
| Botpress | Root-domain discovery or specific pages; sitemap crawler recommended | Documents converted to Markdown, “any native styling or images is removed” | Native Tables feature, each row a chunk | Notion, Google Drive, Slack, Jira, Zendesk, webhooks, Files API | Schedulable re-crawl, including daily |
| Voiceflow | Web pages, sitemap with include/exclude patterns | No, text only | .csv and .xlsx up to 10 MB, “each row is a chunk” | Zendesk, Shopify, Files API | Never, daily, weekly or monthly |
| CustomGPT | Direct sitemap, WordPress, Wix, YouTube | PDF and 1,400+ document formats listed | Not documented | HubSpot, Zendesk, Confluence, Shopify, Google Drive, Notion, SharePoint (Premium+) | Auto-sync from Premium upward |
| RAGFlow | Sitemap ingestion added 10 September 2026 | Yes. DeepDoc does layout analysis, OCR and table recognition, on CPU in 1.0 | Yes, explicit table recognition | Google BigQuery with incremental sync, added 2 July 2026 | Per-dataset |
| Dify | Website sync; Notion sync | Plugin-based; images auto-extracted from DOCX and XLSX under 2 MB, PDF images need pipeline plugins | Chunking templates | Notion, external knowledge base API | On demand |
| AnythingLLM | Local directory sync, Git and other sources via tooling | Multiple document pipelines, collector service | Spreadsheet support | Extensible collector and API | Directory re-sync |
| Open WebUI | Add webpage, sync local directory; remote sources via the oikb tool | Yes, 8 extraction engines including Tika, Docling, Mistral OCR, MinerU, PaddleOCR | Engine-dependent | Native Google Drive and OneDrive/SharePoint | Incremental SHA-256 sync |
Three conflicts worth flagging. Tidio’s own help centre says “Scan priority pages” is limited to 100 pages and 500 URLs overall; a competitor review from eesel.ai states the scan “stops at 60 pages.” The vendor’s documentation wins, but the disagreement exists. Open WebUI’s documentation page lists 13 vector databases while its GitHub README says 9; treat the docs page as newer and the README as out of date.
Side by side: AI chatbots trained on your website and PDFs
| Platform | Billing unit | Entry price, published 16 Oct 2026 | Corpus limits | Best fit |
|---|---|---|---|---|
| Intercom Fin | Per outcome | $0.99 per resolution; $0.99 procedure handoff and qualification; $9.99 disqualification | Not published as a document cap | Teams already on Intercom |
| Zendesk AI agents | Per agent + per automated resolution | $19 / $55 / $115 per agent per month annual; $1.50 committed, $2.00 pay-as-you-go resolutions | 5 to 15 resolutions per agent per month by plan, max 10,000 a year | Zendesk shops |
| Tidio Lyro | Per Lyro conversation | Lyro standalone from $32.50/mo; Growth from $49.17/mo; Plus from $300/mo + usage | 100 pages per crawl, 500 URLs, 10 PDFs at a time, 10,000 Q&A rows | Ecommerce and small stores |
| Chatbase | Per message credit | Hobby $40/mo, Standard $150/mo, Pro $500/mo (list); $32 / $120 / $400 billed annually | 10 MB per file (20 MB on Pro), 700 to 15,000 credits | Non-technical teams |
| Botpress | Per conversation + AI usage | Free 100 conversations; Plus $150/mo annual; Team $750/mo annual | 100 MB to 5 GB vector storage, 1K to 500K table rows | Technical teams wanting control |
| Voiceflow | Plan fee + credits + seats | Four plans, prices only viewable inside the product | Per-document plan limits, not published | Agent builders with budget |
| CustomGPT | Per agent + per query | Standard $99/mo, Premium $499/mo (list); $89 / $449 annual | 5,000 documents per agent on Standard, 20,000 on Premium | Branded assistants at volume |
| RAGFlow | Self-hosted, free software | Open source; cloud option exists | Dataset-based, engineering-controlled | Maximum citation control |
| Dify | Per workspace per year | Sandbox free; Professional $590/yr; Team $1,590/yr | 50 to 1,000 knowledge documents, 50 MB to 20 GB | Self-hosting with a gentle start |
| AnythingLLM | Self-host free; cloud per month | Basic $50/mo, Pro $99/mo, Enterprise custom | Engineering-controlled; MIT licensed | Data that cannot leave the building |
| Open WebUI | Self-host, no licence fee | $0 software, you pay for models and hosting | Engineering-controlled | Advanced internal search |
What it actually costs: normalising five pricing models
The mistake almost everyone makes is comparing entry prices. Entry price tells you almost nothing, because these vendors sell five completely different units.
Here is the workload I used, so you can change any assumption and redo the arithmetic:
- Five human agents, 5,000 inbound contacts a month
- The bot handles 2,000 of them autonomously; 800 bot conversations escalate to a human
- Each bot conversation averages four message exchanges, so 8,000 generated answers a month
- Corpus: a 250-page website, 400 PDFs totalling roughly 3,000 pages, plus a Zendesk Help Center
Intercom Fin. 2,000 outcomes × $0.99 = $1,980/month, and escalated conversations are explicitly not billed, so the 800 handoffs cost nothing. Fin Voice, if you take it, runs at $1.99 per voice outcome and is sold through sales rather than the pricing page. Intercom’s seat fees sit on top of this. Per-seat rates for Essential, Advanced and Expert are not confirmed here; check them before calculating the full bill.
Zendesk AI agents. On Suite Team at $55 per agent per month, five seats is $275 and the plan includes 5 automated resolutions per agent per month, so 50 included. You need 2,000, leaving 1,950 at the $1.50 committed rate = $2,925, for $3,200/month.
On Suite Professional at $115 per seat the subscription is $575 and the inclusion rises to 10 per agent, or 100, leaving 1,900 × $1.50 = $2,850, for $3,425/month. One ceiling to note: Zendesk caps allocated automated resolutions at 10,000 a year across the whole account.
Tidio Lyro. Growth starts at $49.17/month billed annually and covers 250 billable human conversations. Lyro quota is priced separately and tiered by volume; the public pricing page’s calculator did not expose per-tier Lyro prices in static text, so treat $49.17 as the platform floor, not your total.
Chatbase. Credit cost depends on model: most flagship models consume 2 credits per response, GPT-5.5 consumes 4, Claude 5.5 Opus 5, and “all other models” consume 1. At 2 credits per answer, 8,000 answers needs 16,000 credits, so Pro’s included 15,000 leaves you a top-up and lands at about $500/month. Run a cheaper one-credit model instead and Standard’s 4,000 credits covers half the traffic.
Botpress. Team at $750/month billed annually includes 1,500 conversations and $150 of AI usage. 500 more conversations at $50 per 100 = $250. Total $1,000/month before AI usage exceeds the included $150.
CustomGPT. Premium at $449/month billed annually includes 5,000 queries, which comfortably covers 2,000 conversations. About $449/month.
Dify Cloud. Professional at $590/year works out to $49.17/month and includes 5,000 message credits, 500 knowledge documents and 5 GB of storage. 8,000 answers exceeds 5,000 credits, so Team at $1,590/year ($132.50/month, 10,000 credits) is the realistic fit.
Self-hosted RAGFlow, AnythingLLM or Open WebUI. Software cost is zero. What you actually pay for is tokens and hardware, and that varies too much for me to give you a defensible figure without knowing your model and your traffic.
How to deploy one that survives contact with customers
- Pick the abstention behaviour before you pick the model. In RAGFlow, fill in the Empty response field. In Voiceflow, set a scripted Global no-match. In Chatbase, write the constraint into the instructions and test whether the model actually honours it. This decision is worth more than any model choice.
- Audit your PDF corpus for a text layer before uploading anything. Try selecting and copying a sentence from a random page of a typical document. Anything that yields nothing is invisible to Chatbase and Voiceflow, and half-visible to Tidio.
- Turn on re-sync before you go live, not after. Chatbase auto-resyncs weekly and only re-embeds changed pages. Voiceflow offers never, daily, weekly or monthly. Intercom’s web sync is weekly by design. Botpress lets you schedule re-crawls. A bot trained on a stale snapshot is worse than no bot.
- Test the twenty questions you know the answer to. Take the questions support gets most often, write the correct answer from your own documentation, and check the bot against them weekly. Intercom ships an answer debugger, Tidio ships a Suggestions queue of real questions Lyro could not answer, Zendesk’s Quality Assurance flags knowledge gaps.
- Confirm escalation is not billed and not ignored. Intercom does not charge for escalated conversations. Zendesk verifies each resolution with an LLM and does not count ones that fail. Check what your vendor does, because an aggressive bot that never escalates is a cost problem and a liability problem at once.
- Put the disclosure in the first message. One line. EU AI Act Article 50(5) requires it at the latest at the first interaction, in a clear and distinguishable form, conforming to accessibility requirements.
What breaks: failure modes and gotchas
Scanned PDFs are invisible. Already covered, and it is the single most common reason a deployment underperforms. Fix it with OCR before upload, not after.
Re-sync wipes manual edits. Tidio’s documentation warns that when you re-sync a website, Lyro “removes the associated Q&A pairs and replaces them with new ones,” so any manual edits to URL-derived answers are lost. If you curate Lyro’s answers, you are maintaining a patch that the next crawl will delete.
Embedded PDF content produces no citation. Intercom’s own words: uploaded documents “are a private source so your customers will not see any link references to them in Fin’s answers,” and that content is “only there to be used by Fin to improve answer quality and coverage.” Same applies to snippets. Your customers get a confident answer with nothing to click.
Tidio’s AI stops dead when the quota runs out. Every Tidio account starts with 50 Lyro conversations, once. Lyro “stops working” until you pay. Chatbase behaves similarly: when credits run out, the agent displays “This AI Agent is currently unavailable.”
Voiceflow imports PDFs but strips everything else. Documents are “converted to markdown” and “any native styling or images is removed.” If your documentation’s meaning lives in screenshots, that content is gone.
RAGFlow locks your embedding model. Once you have parsed a file with an embedding model, you cannot change it for that dataset, because all files must share one vector space. Choose carefully; re-parsing a large corpus is not free.
Open WebUI’s default retrieval mode will quietly hide your content. Native function calling is now the default, and in that mode attached knowledge is not auto-injected. The model must choose to call the knowledge tools. Open WebUI’s own troubleshooting guide suggests adding a system prompt telling the model to search before answering.
Document-extraction defaults bite in self-hosted stacks. Open WebUI’s docs note that the default pypdf extractor can return empty content for image-based PDFs, and recommend Apache Tika or Docling instead.
Legal and privacy: GDPR, residency and the AI Act disclosure duty
EU AI Act Article 50 is now live. Article 113 of the Regulation sets the application date at 2 August 2026, and the European Commission’s own summary confirms “the transparency rules of the AI Act will come into effect in August 2026.”
Article 50(1) requires providers to ensure that AI systems intended to interact directly with natural persons inform those people they are interacting with an AI system, unless it would be obvious to someone reasonably well-informed, observant and circumspect given the circumstances. Article 50(5) requires that information at the latest at the time of first interaction, in a clear and distinguishable manner, meeting accessibility requirements. The rest of Article 50 is mostly about deepfakes and public-interest text, which does not apply to a support bot.
Data residency. Intercom’s privacy policy is unusually explicit: “The Intercom Services and Sites, and our messenger domains are provided and hosted in the United States.” Transfers rely on the EU-US, UK-US and Swiss-US Data Privacy Frameworks and on Standard Contractual Clauses, and the service-provider list explicitly names “generative artificial intelligence” vendors. Group companies sit in Dublin, London, Berlin, Sydney and San Francisco. If a controller requirement says EU-only processing, that is a conversation to have before you embed, not after.
Training on your data. Chatbase states it twice, on its pricing page and on its security page: data connected to Chatbase “isn’t used to train models” and “your data is never used to train other models.” It carries SOC 2 Type II, GDPR compliance and HIPAA-eligible Enterprise workspaces with a BAA and zero data retention. CustomGPT lists a data processing agreement as an Enterprise-only feature, which is worth noting. Ada claims “zero data retention policies with LLM providers,” alongside SOC 2, HIPAA, GDPR, PCI DSS and AIUC-1.
Open source means you can keep the data home. AnythingLLM is MIT licensed and runs on your own hardware with your own model key. RAGFlow, Dify and Open WebUI are all self-hostable, which is the only reliable way to guarantee that customer content never leaves your network.
This section provides general legal information. If Article 50, GDPR processor terms or a DPIA apply to you, take the vendor’s data processing agreement to someone who can advise on your specific obligations.
The bottom line
If you cannot deploy anything today, pick the platform whose failure mode you can live with, not the one with the prettiest demo.
Choose RAGFlow if the single most important property is that the bot refuses to answer outside your corpus, and you have someone who can run Docker with 16 GB of RAM. Choose Chatbase if you want citations and abstention managed by someone else and your PDFs all have selectable text. Choose Tidio Lyro or Intercom Fin if you want a support platform first and a chatbot second.
Run the test in step four of the deployment list before you sign anything. Ask each shortlisted platform the twenty questions you already know the answer to, and read the answers against your own documentation. If a bot cannot survive that test, it will not survive a customer.
FAQ
Which AI chatbot shows sources with every answer?
Dify documents citations as on by default in chatflows, with a toggle to switch them off. Voiceflow ships a “Show source URL(s)” setting that its documentation explicitly recommends for help centres and documentation. CustomGPT lists “Citations and sources” as included on every plan and claims every response links to its source, but that is a vendor claim with no published benchmark. Botpress exposes citations as a template variable you must render yourself. Chatbase documents no customer-facing citation feature at all.
What happens if a customer asks something that is not in the knowledge base?
It depends entirely on the vendor. RAGFlow returns your fixed “Empty response” string, and its documentation warns that leaving that blank “may give rise to hallucinations.” Voiceflow returns your scripted or generated “Global no-match” reply. Intercom Fin hands over to a human, and the escalated conversation is not billed. Botpress generates a response unless you build the documented Transition Card. Chatbase advises you to write the constraint into the prompt and leaves enforcement to the model.
Can these chatbots read scanned PDFs?
Some can and one vendor says plainly that it cannot. Chatbase: “Scanned PDFs with no text layer cannot be trained on.” Voiceflow: “Only text content is imported.” Tidio reads images inside PDFs even though it does not import them. RAGFlow’s DeepDoc performs layout analysis, OCR and table recognition, on CPU in the 1.0 open-source release. Open WebUI offers eight extraction engines, among them Tika, Docling, Mistral OCR, MinerU and PaddleOCR.
How much does an AI chatbot trained on my website cost per month?
For 2,000 bot conversations a month over a 250-page site and 400 PDFs, the AI component runs from about $49 a month on Dify Cloud Professional to roughly $3,425 a month on Zendesk Suite Professional with committed automated resolutions. Intercom Fin lands near $1,980 at $0.99 per outcome, CustomGPT near $449, Chatbase near $500, and Botpress near $1,000. Self-hosted options cost nothing in licence but shift the cost to models and hardware.
Which platform is cheapest for a five-person support team?
On published list prices, Tidio Growth at $49.17 a month billed annually and AnythingLLM Cloud Basic at $50 a month are the lowest floors, but both sit on separate AI allowances. Zendesk Support Team is $19 per agent per month billed annually with AI billed per resolution. Chatbase Standard is $120 a month billed annually. The cheapest option genuinely depends on whether your unit is messages, conversations, resolutions or pages.
Do I need a developer to set up one of these?
Not for Chatbase, Tidio, Intercom or Zendesk. Chatbase’s FAQ states it needs no coding and offers copy-paste embed codes; Tidio has a no-code dashboard and a testing Playground. Botpress and Voiceflow use visual builders but expect comfort with tools and guardrails. Dify, RAGFlow, AnythingLLM and Open WebUI are self-hostable and expect at least one person who can run Docker; RAGFlow specifies four CPU cores, 16 GB of RAM and 50 GB of disk.
Do these vendors train their models on my content?
Chatbase says no, twice, on its pricing and security pages, and adds SOC 2 Type II, GDPR and HIPAA-eligible Enterprise support. AnythingLLM is MIT licensed and self-hosted, so content can stay on your own hardware. Dify Cloud and self-hosted Dify differ, and CustomGPT lists a data processing agreement as Enterprise-only. Intercom lists generative AI vendors among its service providers, so read the DPA rather than assume.
Do I have to tell customers they are talking to an AI?
In the EU, yes. Article 50(1) of the AI Act requires providers of systems intended to interact directly with people to inform them, unless it is obvious to a reasonably well-informed, observant and circumspect person. Article 50(5) requires that information at the latest at the time of the first interaction, clearly and distinguishably, meeting accessibility requirements. Article 50 has applied since 2 August 2026.
RAGFlow or Dify, which is better for self-hosted document chat?
Dify is the gentler on-ramp: two CPU cores and 4 GB of RAM, a visual workflow canvas, and citations on by default. RAGFlow is heavier and more controllable: four CPU cores, 16 GB of RAM, 50 GB of disk, CPU-based layout analysis, OCR and table recognition, and a documented fixed response for unanswered questions. Dify also locks your knowledge base to a plan: 500 documents and 5 GB on Professional.
Are the prices on other comparison articles reliable?
Mixed. FastBots’ February 2026 guide lists Tidio Starter at $29 a month billed annually and Botpress Plus at roughly $89 a month. Tidio’s own pricing page showed $24.17 a month billed annually and Botpress’s own page showed $150 a month billed annually on 16 October 2026, so those figures are off by roughly 22% and 69%. Chatloom’s guide lists Chatbase from $19 a month; Chatbase’s own page shows $40 monthly or $32 billed annually. Always check the vendor’s page.
Sources
- Intercom Pricing: Fin outcome prices of $0.99 and $9.99, Fin Voice at $1.99, Pro and Copilot add-on pricing
- Fin AI Agent explained and Fin AI Agent outcomes (Intercom Help) capability list, human handoff triggers, escalation not billed
- Add your content for Fin AI Agent and Knowledge sources (Intercom Help) weekly web sync, PDF and snippet privacy, source-type matrix
- Debug Fin AI Agent answers: Intercom Help: staff-side answer debugger and snippet behaviour
- Intercom Privacy Policy: US hosting, DPF and SCC transfers, generative AI in the service-provider list
- Zendesk Pricing: $19 / $55 / $115 per agent per month annual, $1.50 committed and $2.00 pay-as-you-go resolutions
- About automated resolutions and About AI agents with agentic AI (Zendesk Help Center) plan allotments, LLM verification, escalation triggers
- Advanced AI Agents for CX, Zendesk Copilot and Zendesk Knowledge: external sources including Google Drive and PDFs; Copilot priced via account executive
- Tidio Pricing: Starter $24.17, Growth from $49.17, Plus from $300, Lyro standalone from $32.50, guidance and upload limits
- Data sources (Lyro’s knowledge base) Tidio Help: 100-page crawl cap, 500 URLs, 10 PDFs, source links, re-sync behaviour
- An honest review of Tidio’s Lyro AI in 2026: eesel AI: reviewer claim of a 60-page crawl cap, conflicting with Tidio’s own documentation
- Chatbase Pricing and Chatbase Security: Hobby $40 / $32 annual through Pro $500 / $400, credits, file sizes, “not used to train models”
- Data sources and Best Practices (Chatbase Docs) selectable-text PDF restriction, auto-resync cadence, prompt-based abstention
- Chatbase FAQ: credit consumption per model, exhaustion behaviour, data storage statement
- Botpress Pricing: Free, Plus $150/mo annual, Team $750/mo annual, conversation definition, $40/mo storage add-on
- Knowledge Bases and Add sources to Knowledge Base (Botpress Docs) no-answer Transition Card recipe, crawl modes, re-crawl schedules
- Knowledge Agent and Knowledge Base Best Practices: citation variable; Markdown conversion stripping images
- Retention Period: Botpress Docs: logs 30 days, conversations and messages 90 days
- Importing data sources and Querying the knowledge base (Voiceflow Docs) data types, 10 MB text-only PDF import, refresh rates, Show source URL(s)
- Behaviour and Billing overview (Voiceflow Docs) global no-match behaviour; four plans with pricing only viewable in-app
- Global prompt: Voiceflow Docs: documented anti-hallucination guardrail examples
- CustomGPT Pricing: Standard $99 / $89, Premium $499 / $449, query and document limits, anti-hallucination and citation claims
- Dify Pricing and Knowledge Retrieval node: Sandbox, Professional $590/yr, Team $1,590/yr; citations on by default
- Upload Local Files and self-hosting requirements (Dify Docs) 15 MB and 50 MB upload limits, image extraction rules, 2 CPU cores
- langgenius/dify on GitHub: Dify Open Source License based on Apache 2.0 with additional conditions
- infiniflow/ragflow on GitHub: release dates, grounded citations claim, DeepDoc CPU inference, system requirements
- RAGFlow quickstart: Empty response abstention setting, embedding model lock, chunk intervention
- AnythingLLM on GitHub: MIT licence, source citations, telemetry disclosure, supported providers
- AnythingLLM Cloud pricing and AnythingLLM Docs: Basic $50/mo, Pro $99/mo; chunk retrieval and refusal behaviour
- Open WebUI Knowledge Bases: vector databases, extraction engines, retrieval modes, agentic tools
- Open WebUI Document Extraction and RAG troubleshooting: scanned PDFs and OCR engines; pypdf limits and native function calling behaviour
- open-webui on GitHub: licence and branding requirement, 9 vector databases as stated in the README
- ChatPDF, AskYourPDF pricing and Humata pricing: PDF-first tools: citations, OCR tiers and per-page pricing
- Ada: no public pricing, zero data retention with LLM providers, compliance badges (vendor claims)
- Forethought pricing, Glean and Salesforce Agentforce pricing: enterprise shortlist: platform fee plus outcome pricing, Flex Credits, per-user licensing
- Regulation (EU) 2024/1689, Article 50 and Article 113 (EU AI Act explorer) disclosure duty and application dates
- AI Act: European Commission: transparency rules effective August 2026, AI Omnibus timeline
- CiteShade: Citation Laundering in Multi-Source RAG (arXiv:2609.15660): wrong-answer rate 0.01 to 0.68; preprint, submitted 14 September 2026
- Configurable Clinical Information Extraction with Agentic RAG (arXiv:2606.19602): 7,326 clinician judgments, 96.5% acceptance; preprint, submitted 17 June 2026
- Invisible Ink, Visible Lies (arXiv:2610.04860): watermark-induced hallucination, NeurIPS 2026
- FastBots, Chatloom and WPBeginner: the three ranking competitor comparisons and their February–July 2026 figures
