Concepts
How FlexyAgents ingests content, chunks and embeds it, retrieves relevant passages, and steers the model to stay grounded in what you approved.
Retrieval-augmented generation means the model receives excerpts from your corpus before it answers. FlexyAgents handles ingestion pipelines, vector search, and ranking so operators focus on content quality and scope—not low-level ML plumbing.
When retrieval finds nothing useful, well-tuned agents should admit uncertainty and offer human paths instead of inventing facts.
Uploads cover PDFs, Office documents, HTML exports, images, audio, video, and other supported binaries. Text extraction may combine parsers, OCR, and optional Gemini vision or transcription when configured. Connectors sync SaaS sources on a schedule. Crawlers pull public URLs with depth and politeness controls and can optionally process linked images and media.
Structured Q&A rows complement long documents for answers that must be verbatim—policies, legal blurbs, or SKU facts.
Documents split into chunks sized for embedding models; boundaries try to respect headings and tables but imperfect sources still benefit from editorial cleanup.
Queries embed into the same space as chunks; top passages are ranked and injected into the prompt with instructions to cite or paraphrase faithfully.
Product behavior encourages citing sources customers recognize (article titles, URLs) when your template requests it. Refusal templates reduce hallucination risk for high-stakes topics.
Analytics on “unknown” or low-confidence turns helps prioritize knowledge gaps without reading every transcript.
Continue exploring
Platform features · HTTP API & webhooks · Quickstart · Contact
Build on your stack
Start a trial or talk to us about rollout, governance, and enterprise requirements.