Core Concepts
Knowledge Base
Ground your agent in your real data. Talknex uses pgvector RAG to pull relevant context from your documents before every LLM call — eliminating hallucinations.
How RAG works
When the caller says something, the system embeds their utterance using OpenAI's text-embedding-3-small model (1536 dimensions) and performs a cosine similarity search against your org's vector store. The top-K matching chunks are injected into the LLM context before generation.
# Simplified flow
caller utterance → embed (1536-dim) → pgvector cosine search
→ top 3 chunks injected into system prompt
→ GPT-4o generates grounded response
Ingestion sources
File upload
Drag and drop PDFs, DOCX, TXT, or Markdown files. Maximum 50MB per file. Multiple files can be uploaded in bulk. The system parses, chunks, and embeds in background.
PDF · DOCX · TXT · MD · CSVURL crawl
Provide a root URL and Talknex crawls up to 500 pages. Good for product documentation sites, help centers, and FAQs. Recrawls can be scheduled daily or weekly.
Any publicly accessible URLAPI ingestion
POST raw text chunks directly to the Knowledge API. Use this for dynamic content that changes frequently — your backend pushes updates in real-time.
REST API · WebhooksChunking strategy
Documents are split into chunks before embedding. The default strategy is semantic chunking (splits at paragraph and sentence boundaries), which produces higher-quality retrieval than fixed-size character splitting.
Managing collections
Knowledge sources are organized into Collections. Each collection can be attached to one or more agents. A collection can contain mixed sources (files + URLs).
Plan limits
| Plan | Collections | Documents | Crawl pages | Re-index freq |
|---|---|---|---|---|
| Free | 1 | 10 | 50 | Manual |
| Pro | Unlimited | 500 | 500 | Daily |
| Enterprise | Unlimited | Unlimited | Unlimited | Real-time |