TileScore

    AI Consultant Data Import Guide

    Formats, structure and requirements for TileScore AI Consultant (Nanobot) training data

    1. Architecture Overview

    TileScore AI Consultant (Nanobot) uses RAG (Retrieval-Augmented Generation) architecture to provide accurate answers based on client data. The system combines vector search with full-text search for maximum relevance.

    • Embedding Engine — DashScope text-embedding-v3, dimension 1024
    • Vector search — pgvector (cosine similarity), threshold 0.3
    • FTS — PostgreSQL tsvector with Russian morphology
    • Cascading search — current tile → children → parents → guest tiles
    • Chunking — 500 chars, overlap 100

    2. Data Import Formats

    ChannelFormatsPurposeLimit
    KB DocumentsPDF, TXT, MD, JSONTexts, instructions10 MB, 100 files
    KB CatalogXLSX, XLS, CSVProducts, services, prices50 000 rows
    URL AutofillURLWebsite parsing5 sites, 20 pages
    Q&A TrainingUIQ&A pairsUnlimited

    3. Knowledge Base Documents

    Supported formats

    FormatExtensionNotes
    Text.txtPlain UTF-8 text. Recommended.
    Markdown.mdHeaders used as chunk delimiters.
    PDF.pdfText-based only. Scans NOT supported.
    JSON.jsonValidated and formatted.

    Requirements

    • Min 10 chars, max 500,000
    • UTF-8 encoding
    • PDF: text-based only (not scans)
    • Built-in spam check

    TXT Example

    # О компании

Компания «Пример» основана в 2010 году.

# Услуги

1. Консультирование — индивидуальный подход
2. Аудит — комплексная проверка

# Контакты
Телефон: +7 (999) 123-45-67

    4. Catalog Data

    Required fields

    • name — Product/service name
    • code — Unique code/SKU

    Optional fields

    • description — Detailed description
    • category — Category/group
    • price — Price in any format
    • unit — Unit of measurement
    • metadata — Additional properties (JSON)

    CSV Example

    Код;Наименование;Категория;Описание;Цена
PN-001;Перчатки нитриловые;СИЗ;Размер M, синие;450
PN-002;Перчатки латексные;СИЗ;Размер L, белые;380
RS-010;Респиратор FFP2;СИЗ;Одноразовый;120
    CSV delimiter — semicolon (;) or comma (,). Auto-detected. UTF-8 encoding.

    Catalog search (3 levels)

    • Exact code match (rank 100)
    • Full-text search (FTS) with Russian morphology
    • ILIKE fallback by keywords (≥4 chars)

    5. URL Autofill

    • Up to 5 websites
    • Up to 20 subpages per site
    • Text content extracted automatically
    • Navigation and service blocks filtered

    6. Training Q&A

    CategoryPurpose
    PassportBasic company info
    Q&ACommon client questions
    FunnelSales funnel questions
    CorrectionAI answer corrections (highest priority)

    7. Limits

    ParameterValue
    Max file size10 MB
    Max docs per tile100
    Min content length10 chars
    Max content length500,000 chars
    EncodingUTF-8
    Document formatsPDF, TXT, MD, JSON
    Catalog formatsXLSX, XLS, CSV
    URL Autofill5 sites, 20 subpages
    PDF OCRNot supported

    8. Data Processing

    Stage 1: Upload & Validation

    Format, size, encoding check. Whitespace normalization. Spam and duplicate check.

    Stage 2: Chunking

    Chunk size: 500 chars. Overlap: 100 chars. Split at sentence boundaries.

    Stage 3: Embedding

    DashScope text-embedding-v3, 1024 dims → pgvector

    Stage 4: Indexing

    FTS (tsvector) + vector index + catalog code index

    File → Validation → Chunking (500/100) → Embedding (1024d) → pgvector + FTS

    9. Recommendations

    Documents

    • Use TXT or MD — most reliable formats
    • One document = one topic
    • Avoid tables in text files — use catalog

    Catalogs

    • First row — column headers
    • Code must be unique
    • Description — more detail = better search

    Q&A

    • Phrase questions like customers would
    • Use "Correction" to fix AI errors