Bot & AI
Upload your PDFs, FAQ docs, product catalog, policies. Your bot retrieves the right context for every question and answers from your actual content — no hallucinations, no guesswork.
Per document
PDF / DOCX / TXT
Documents
Per knowledge base
Answer accuracy
Grounded answers
Retrieval time
Vector search
PDFs, Word docs, Markdown, plain text, even web pages. Edesy parses, chunks, and embeds each document automatically. No file format conversion required.
Every chunk gets embedded and stored in Pinecone. When a customer asks a question, semantic search finds the 3–5 most relevant chunks in under 500ms.
Configure per bot: inject top-K chunks into the system prompt automatically, or run a fresh retrieval per message, or both. Tune precision vs cost per bot.
The LLM is instructed to answer only from retrieved context. If the answer isn't in your docs, the bot says so explicitly rather than making something up.
Point the bot at your website URL. Edesy crawls and indexes the pages. Re-crawl on a schedule so your bot stays in sync with site updates.
Each knowledge base lives in its own Pinecone namespace. Your data never bleeds across customers. SOC 2 controls on the underlying infrastructure.
PDFs of your product brochures, your help center FAQ, your company handbook, your training manuals. Edesy supports any size — common stores have 50–500 documents.
Each document is split into semantic chunks (~500 tokens each), each chunk gets a vector embedding via OpenAI's text-embedding model, and the vectors are stored in Pinecone with metadata pointing back to the source document.
A customer sends a message like 'what's your return policy for opened items?'. The bot embeds the question and runs a similarity search across the knowledge base.
The 3–5 most relevant chunks (e.g. excerpts from your returns policy doc) are pulled and inserted into the LLM's context window alongside the system prompt and conversation history.
The LLM (Gemini, GPT-4, or Claude) reads the question, the retrieved context, and the system prompt, then composes an answer using only the retrieved facts. If the docs don't contain the answer, the bot says 'I don't have that info — let me connect you with a human'.
| Feature | Edesy | Typical WhatsApp platform |
|---|---|---|
| Upload custom documents | Limited | |
| PDF + DOCX + URL crawling | PDF only | |
| Vector search (Pinecone) | ||
| Per-bot KB configuration | ||
| Pre-inject + query-time retrieval | One mode | |
| Grounding to prevent hallucinations | ||
| Per-workspace data isolation | Shared | |
| Document processing speed | <2 min for 100 pages | Hours |
Policy documents, coverage details, claim procedures. Bot answers 'is roof leak covered under my home policy?' from the actual policy PDF.
Agent time on policy questions down 70%
Course syllabi, batch schedules, refund policies. Bot answers parent questions accurately and routes the rest to admissions.
Lead-to-counselor conversation rate up 40%
Doctor profiles, specialty areas, insurance accepted, FAQs about visits. Bot books appointments and answers logistics questions.
Front-desk call volume reduced by half
Product ingredient lists, skin-type guides, usage instructions. Bot answers 'is this safe for sensitive skin?' from product docs.
Pre-purchase support response time <1 min
Property amenities, room types, local guides, FAQ. Bot answers booking questions across 20 properties from one knowledge base.
Direct booking rate up 18% vs OTA-only
Help center articles, API docs, pricing pages. Bot handles tier-1 support entirely, escalates only edge cases.
Tier-1 support ticket volume reduced 65%
A WhatsApp bot built on a generic LLM (GPT-4, Gemini, Claude) can hold a fluent conversation, but it doesn't know anything specific about your business. Ask it 'what's your return policy?' and it'll make up a plausible-sounding answer based on what it thinks 'typical' return policies look like. For 70% of questions that answer is close enough to be undetectable — and for the 30% where it's wrong, you've just damaged the customer relationship in a way you can't see.
Retrieval-augmented generation (RAG) solves this by separating the language ability from the knowledge. The LLM still does the writing — handling tone, grammar, multilingual translation, conversational flow — but the *facts* come from your own documents. Every answer is grounded in retrieved chunks of YOUR content. If the answer isn't in your docs, the bot says so transparently instead of hallucinating.
The infrastructure that makes this possible is vector embeddings + similarity search. When you upload a document, Edesy splits it into ~500-token chunks and generates a high-dimensional embedding (1,536 dimensions, typically) for each chunk via OpenAI's text-embedding model. Those vectors get stored in Pinecone, a managed vector database. When a customer asks a question, the question itself is embedded and Pinecone returns the chunks whose embeddings are most similar — usually within 500ms. Those chunks become the 'context' that the LLM uses to answer.
The practical impact: instead of having to hand-write every possible FAQ into your bot's system prompt (which doesn't scale past a few hundred lines), you can dump your existing knowledge — help center articles, product PDFs, training manuals, policy documents — and have the bot answer accurately across all of them. A typical Edesy customer loads 50–500 documents and replaces 60–80% of their tier-1 support volume within a month.
Upload your existing documents, point the bot at your help center, and watch it become an expert on your business. Free workspace, no credit card.