v2026.3.6

Release date: March 24, 2026

---

✨ New Features

Knowledge Base for AI Agent

Each company can now register a Knowledge Base directly in the AI Agent. Registered documents are automatically made available as tools for the AI, which consults them during service interactions before any external lookup.

What is it for?

• Precise answers – The AI consults your company's own information: hours, addresses, policies, FAQs, price tables and more
• Token reduction – By using internal data instead of reasoning from generic knowledge, token consumption drops significantly
• No extra setup – Each registered document automatically becomes a tool. No additional action is required

How does the search work?

The search is adaptive based on document size:

• Short documents (up to 3 chunks) → full content returned directly
• Long documents (more than 3 chunks) → semantic RAG search returns only the most relevant excerpts for the customer's question

Internally, a hybrid search is used combining vector embeddings (OpenAI text-embedding-3-small) with keyword search (tsvector), supporting Portuguese, English and Spanish.

How to use?

1. In the left sidebar, go to AI Agent
2. Click Edit on the desired agent
3. In the top bar, click the Knowledge Base tab
4. Click Add document
5. Fill in the fields:
   • Title – Document name (e.g. Business Hours)
   • Description – Summary of what the document contains. The more specific, the better the AI knows when to call it (e.g. Contains business hours, holidays, phone and WhatsApp for scheduling)
   • Language – Primary content language
   • Content – Full text, with Markdown support
6. Click Save

Tip

The description is the most important field. Use words the customer would use when asking. Example: instead of "General information", write "Contains address, hours, accepted insurance plans and consultation fees".

Organize by topic

Prefer one document per topic (e.g. Hours, Address, Cancellation Policy) instead of one large document with everything. Focused documents produce more accurate answers.

---

🐛 Bug Fixes

Deduplication of repeated chunks in context

When the AI called the same KB tool in consecutive turns, the history accumulated the same content multiple times, unnecessarily inflating the context. Now only the most recent call for each KB tool is kept in the history sent to the model.

Corrupted text when editing long documents

When opening a multi-chunk document for editing, the text appeared with repeated or mid-sentence cuts. Content reconstruction now detects and removes the overlap between consecutive chunks, returning the original text without duplications.

---

🎯 Benefits

• ✅ Faster and more accurate answers – AI consults internal company data without relying on generic reasoning
• ✅ Reduced token consumption – Relevant, non-repetitive content sent to the model
• ✅ Zero configuration – Documents become tools automatically
• ✅ Multilingual support – Keyword search in Portuguese, English and Spanish
• ✅ Adaptive RAG – Short documents returned in full; long ones go through semantic search
• ✅ Markdown editor – Content can be formatted with bold, lists, headings, etc.