Upload your FAQs, policies, product docs, and service menus. They're chunked, embedded, and retrieved by semantic search on every customer message. Your assistant always answers from your content — never from thin air.
Paragraph-aware boundaries and token overlap preserve context edge-to-edge.
Concept-driven retrieval matches intent rather than just literal keyword hooks.
Documents are immediately indexed and available on WhatsApp and Voice channels.
| Filename | Size | Status |
|---|---|---|
| pdfclinical_guidelines.pdf | 1.2 MB | Ready |
| docxservices_pricing.docx | 340 KB | Ready |
| txtreceptionist_faq.txt | 45 KB | Ready |
Upload your FAQs, policies, and service menus once. CareFlow's paragraph-aware chunking and semantic vector indexing ground your AI assistant in real team data.
Files are split into 500-token chunks with paragraph-aware boundaries and 50-token overlap. No sentence is cut in half.
Each chunk is embedded with OpenAI's models. When a customer asks a question, the system finds conceptually similar content.
On every message, the assistant searches your knowledge base. The top 5 relevant chunks are injected into context in milliseconds.
Upload whatever you have. PDFs are parsed with table-aware layout detection. Markdown stays clean. Up to 25MB per file.
Documents are available on WhatsApp the moment processing completes. For voice, publish your assistant to sync.
Set documents as available to all assistants, or lock them to specific ones. Sensitive internal policies stay secure.
Test semantic search across your files. See exactly which chunks match, with similarity scores. Confirm accuracy fast.
Every document shows its status: processing with a progress percentage, ready in green, or failed with an error message.
| Filename | Tokens | Chunks | Status |
|---|---|---|---|
| clinic_faq.pdf | 14,500 | 29 chunks | Ready |
| services_menu.docx | 4,200 | 9 chunks | Processing 72% |
Drag and drop your documents. Set visibility — available to all assistants, or restricted to specific ones. PDFs, Word docs, text files, CSVs, and Markdown are all supported.
Your document flows through a five-stage pipeline: layout-aware extraction, token chunking (500 tokens, 50 overlap), vector embedding, database indexing, and runtime mapping.
Your assistant pulls from this content on every message. Open any document to inspect its chunks. Use the search test tab to verify results with similarity scores.
Upload them. Your assistants will find the right answer, every time — without you writing a single prompt. Semantic search does the work.