Перейти к содержимому

Knowledge base

Company memory for AI agents

Upload policies, contracts, brand guidelines and FAQs — the agent pulls the relevant passages on every request and answers with direct quotes. RAG (Retrieval-Augmented Generation) makes hallucinations rare.

Start free

How it works

Search, not memorization

Documents are split into chunks and indexed. On every request the model receives the ten closest passages — it never tries to read everything.

  • PDF, DOCX, XLSX, PPTX

    Also TXT, Markdown, CSV, JSON, XML. Text is extracted.

  • OpenAI embeddings

    text-embedding-3-small — the industry standard for semantic search quality.

  • Source citations

    The agent's response includes a citations field with the document name, the quoted text and its position. You can also require “answer only from the knowledge base” in the prompt.

Use cases

  • Internal handbook

    Policies, procedures, FAQs — employees ask the agent

    Вручную15 min
    С ИИ-агентом30 sec

    to find an answer in a policy document

  • Customer support

    Product FAQ plus the contract — the agent answers with references

    Вручную20 min
    С ИИ-агентом2 min

    per routine ticket

  • Reviewing 15 contracts

    Upload them, then ask which clauses are non-standard

    Вручную8 h
    С ИИ-агентом30 min

    to audit the batch

Common questions

Free — 2, STANDARD — 5, PRO/ULTRA — 20, TEAM5/10/20 — 100, TEAM50 — 250.
Scans without OCR aren't indexed as text. Run them through recognition (or switch on the OCR toggle) before uploading.
Put this in the system prompt: “Answer strictly from the knowledge base. Give a direct quote for every claim. If the base has no answer, say so.” It cuts hallucinations sharply.
Upload it again. There are no real-time updates: the index is rebuilt at upload.

Upload your first documents

Ten five-page files beat one fifty-page file — search relevance is higher.

Start free