Knowledge Base

Train your agent with documents and content

Overview

The Knowledge Base uses RAG (Retrieval-Augmented Generation) to give your agent access to your business content. Documents are split into chunks, converted to vector embeddings, and stored for semantic search.

Navigate to the Knowledge Base tab in the agent panel. Here you choose which data the agent will use for its responses.

Adding Content

Documents

Upload files in supported formats:

• PDF
• DOCX
• TXT

URLs

Provide web page URLs. Revol will scrape the content and add it to the knowledge base.

Text

Add content directly as text blocks.

How RAG Works

1. Upload

   You upload a document or add content.

2. Chunking

   Content is split into manageable chunks.

3. Embedding

   Each chunk is converted to a vector embedding using the selected embedding model.

4. Storage

   Embeddings are stored in PostgreSQL with pgvector extension.

5. Retrieval

   When a user asks a question, the most similar chunks are retrieved using cosine similarity.

6. Generation

   Retrieved chunks are injected into the LLM prompt as context.

RAG Settings

Open RAG settings via the gear icon (⚙) in the Knowledge Base storage panel. All settings are per-company and auto-saved.

Embedding Model

Choose which model converts your text into vector embeddings:

Model	Provider	Dimensions	Price	Best for
text-embedding-3-small	OpenAI	1536	$0.02/1M tokens	General use, English content
text-embedding-3-large	OpenAI	1536	$0.13/1M tokens	Higher accuracy, English content
BGE-M3	DeepInfra	1024	$0.01/1M tokens	Multilingual content (100+ languages)

Changing the embedding model deletes all existing embeddings for the company. You must re-train all agents after switching. A confirmation dialog will appear before the change is applied.

Chunk Limit

How many text chunks are returned per RAG search (1–20). Default: 5.

Higher values provide more context to the LLM but increase token usage.

Character Limit

Maximum characters per chunk when splitting documents (500–10,000). Default: 1,500.

Smaller chunks give more precise retrieval. Larger chunks preserve more context per result.

Chunk Overlap

Overlap between consecutive chunks (0–40%). Default: 15%.

Overlap ensures important context at chunk boundaries is not lost. Higher overlap creates more chunks and uses more storage.

Similarity Threshold

Minimum cosine similarity score to include a result (0.1–1.0). Default: 0.35.

Lower values return more results (better recall). Higher values return only highly relevant results (better precision). For multilingual content, use lower thresholds (0.3–0.4).

Storage Limits

Plan	Knowledge Docs	Embedding Tokens
Free	10	100,000
Premium	100	1,000,000
Professional	1,000	5,000,000

---

Generate Knowledge Base with Claude Code

If you have project documentation (website pages, docs portal, README files, wiki) and want to turn it into a structured knowledge base for your Revol AI agent, you can use Claude Code to analyze the documentation and generate ready-to-upload TXT files.

Prompt for knowledge base generation

# Instruction: Generate Knowledge Base Files from Documentation

## Context

I use Revol — a platform for building AI sales agents with RAG
(Retrieval-Augmented Generation). My agent answers questions
using a knowledge base: uploaded files are chunked (configurable
size and overlap), embedded via a selectable model (OpenAI
text-embedding-3-small/large or BGE-M3), stored in PostgreSQL
with pgvector, and searched by cosine similarity at inference time.

I need to convert my project documentation into a set of .txt
files optimized for this RAG pipeline — so the agent can find
and cite accurate answers.

## How RAG Chunking Works (important for file structure)

- Each file is split into chunks of ~2000 characters with 20%
  overlap at sentence boundaries
- The file name becomes part of the embedding metadata — use
  descriptive names so the system can identify the source
- Shorter, focused files perform better than one giant file
- Each chunk should be self-contained — a reader should
  understand the chunk without needing the surrounding text
- Structured content (lists, tables as text, clear headers)
  chunks better than long prose paragraphs

## Your Task

### Step 1: Analyze All Documentation

Read every documentation file in the project. For each file,
identify:

1. The topic it covers
2. Key facts, settings, configurations, and values
3. Step-by-step procedures
4. Feature descriptions with specifics (limits, options, formats)
5. Pricing, plans, and quotas (if applicable)
6. Technical details (APIs, parameters, integrations)

### Step 2: Plan the File Structure

Group related content into logical topics. Each file should
cover ONE cohesive topic. Target file sizes:

- Ideal: 2,000–6,000 characters per file (1–3 chunks)
- Maximum: 10,000 characters (5 chunks)
- If a topic is larger, split it into sub-topics

Naming convention: {Topic-Name}.txt
- Use descriptive English names with hyphens
- No numbering prefixes (ordering doesn't matter for RAG)
- Name should hint at the content for metadata

Examples:
- Platform-Overview.txt
- Getting-Started.txt
- Pricing-Plans.txt
- API-Authentication.txt
- Webhook-Integration.txt

### Step 3: Write the Files

For each file, follow these rules:

Content rules:
- Start with a clear topic heading as plain text
- Write in plain text — no Markdown formatting, no HTML,
  no JSX components
- Convert tables to readable text lists or key-value pairs
- Convert step-by-step guides to numbered lists
- Include specific values: numbers, limits, prices, options,
  defaults, formats, URLs
- Each paragraph or section should be understandable on its
  own (self-contained chunks)
- Remove navigation elements, "Next Steps" links, and
  cross-references that don't add informational value
- Keep the original language of the documentation

Optimization for RAG:
- Front-load important information — put the answer before
  the explanation
- Use consistent terminology throughout all files
- Repeat key terms naturally so they appear in multiple
  chunks (improves retrieval)
- For features with settings: list setting name, type,
  default value, options, and description
- For integrations: list provider, auth method, setup steps,
  available tools/functions

What to exclude:
- Screenshots and image references
- UI component markup (tabs, accordions, cards, frames)
- "See also" and navigation links
- Decorative text and marketing fluff
- Duplicate content (don't repeat the same info in
  multiple files)

### Step 4: Create the Output

1. Create a temporary directory (e.g., temp-knowledge-base/)
2. Write all .txt files into it
3. List all files with their sizes and topic descriptions
4. Provide a summary of total files and total size

## Optional: Generate a Workflow Agent Seeder

If the project uses Revol and you also need to create an AI
agent with a workflow to answer questions about this
documentation, generate a Laravel database seeder that:

1. Creates an AiAgent with:
   - Appropriate system prompt for documentation support
   - LLM: gpt-4o-mini, temperature: 0.3 (accuracy)
   - Personality: high clarity (8-9), high formality (6-7),
     low humor (2-3), low emoji (1-2)

2. Builds a Smart Router workflow:
   - Start node with keyword-based edges to expert nodes
   - 3-6 custom "expert" nodes, each covering a topic group
   - Each expert has a specific conversation_goal describing
     its domain
   - Each expert has search_documents + get_company_info tools
   - One fallback edge to a General Assistant node
   - All experts connect to a Formatter node
   - Voice nodes (STT/TTS) inactive by default

3. Keywords on edges should match the natural vocabulary
   users would use when asking about each topic

## Input

My documentation is located at: [PATH OR URL]
Project name: [NAME]
Project description: [BRIEF DESCRIPTION]
Target company_id for seeder: [ID or skip seeder]

This prompt works best with comprehensive documentation portals, product docs, API references, and knowledge bases. The generated .txt files are ready to upload directly to Revol’s File Manager — just drag and drop, then click Train.