What you'll end up with

By the end of this guide you'll have an AI assistant that:

Knows the content of your specific documents
Answers questions with citations back to the source
Is queryable from the dashboard or via a REST API
Costs nothing to set up — and fractions of a cent per question to run

No fine-tuning. No GPUs to manage. No infrastructure setup. Just your documents and a Lexora account.

Why not just use ChatGPT?

General-purpose models like ChatGPT are trained on web data up to a cutoff date. They don't know your internal docs, your product specs, your legal policies, or anything proprietary to your business.

You could paste documents into the context window every time — but that's expensive (you pay for every token, including the document), slow (large pastes add latency), and impractical at scale (one PDF per conversation doesn't work for an assistant handling hundreds of users).

RAG solves all of this. Your documents are indexed once. At query time, only the relevant chunks are retrieved — typically a few hundred tokens — so responses are fast, cheap, and grounded in real content.

Step 1: Create a free account

Go to lexora.network/auth/signin and sign up. No credit card required. You get immediate access to the knowledge base dashboard and a free API key.

Step 2: Create a knowledge base

From the dashboard, open the Knowledge Bases section and create a new KB. Give it a name that reflects what it covers — for example, "Product FAQ", "Employee Handbook", or "Legal Policies".

You can create multiple knowledge bases and switch between them when querying, so it's fine to segment by topic or audience.

Step 3: Upload your documents

Drag in PDFs or TXT files. Lexora's worker network runs the full ingest pipeline automatically:

Text is extracted from your file (including multi-page PDFs)
Content is split into ~1,000-token chunks with overlap to preserve context across boundaries
Each chunk is embedded with BGE-M3 — a multilingual, high-quality embedding model running on CPU workers
Vectors are stored in pgvector and indexed for retrieval

For most documents, this completes in under two minutes. You'll see a chunk count when processing is done — that's how many searchable segments your document produced.

Step 4: Test it in the dashboard

Open the chat interface, select your knowledge base, and ask a question in plain English. The assistant will retrieve the most relevant chunks from your documents and generate an answer that cites the source.

Try a few different questions — specific lookups, summary requests, comparisons across sections. If the answers are off, check whether the relevant content is in your documents; RAG can only surface what's there.

Step 5: Connect via API

Once your knowledge base is working in the dashboard, you can wire it into your own application using your Lexora API key. The API is OpenAI-compatible, so existing integrations require minimal changes.

Here's a minimal working example in Python:

from openai import OpenAI

client = OpenAI(
    api_key="your-lexora-api-key",
    base_url="https://lexora.network/v1"
)

response = client.chat.completions.create(
    model="Qwen/Qwen3-8B",
    messages=[
        {
            "role": "user",
            "content": "What is our return policy for digital products?"
        }
    ],
    # pass your KB ID to enable RAG retrieval
    extra_body={"kb_id": "your-knowledge-base-id"}
)

print(response.choices[0].message.content)

And in TypeScript/Node:

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: "your-lexora-api-key",
  baseURL: "https://lexora.network/v1",
});

const response = await client.chat.completions.create({
  model: "Qwen/Qwen3-8B",
  messages: [
    {
      role: "user",
      content: "Summarize the key points from our Q2 report.",
    },
  ],
  // @ts-expect-error — extra body param
  kb_id: "your-knowledge-base-id",
});

console.log(response.choices[0].message.content);

What it costs to run

Creating and training the knowledge base is free — embeddings, chunking, and vector storage are all included at no charge.

The only cost is the language model generating the final answer. Lexora's in-house Qwen3 8B model costs $0.10 per million tokens. A typical RAG-augmented response is around 400–600 tokens in and 200–400 tokens out — roughly $0.00006 per question. You can run tens of thousands of queries before you hit a dollar.

When to upgrade to a frontier model

Qwen3 8B handles most Q&A, summarization, and lookup tasks well. If you need deeper reasoning — complex multi-document synthesis, nuanced legal interpretation, code generation from specs — you can attach the same knowledge base to a partner model like DeepSeek V4 Pro or Kimi K2.7 Code. Same API, same KB ID, just a different model string.

Use cases this is built for

Customer support bots — answer product questions without human intervention, cite exact docs
Internal help desks — HR policies, IT runbooks, company FAQs available 24/7
Developer documentation assistants — let users query your API docs in natural language
Research tools — upload papers, reports, or meeting transcripts and query across them
Onboarding assistants — new hires can ask the handbook instead of pinging colleagues

The short version

Sign up free. Upload documents. Get a working AI assistant that knows your content, cites its sources, and can be accessed via the OpenAI-compatible API from any stack. Training is free. Inference starts at $0.10/million tokens. No credit card, no GPU, no infrastructure headache.

How to Build a Free AI Assistant Bot with Your Own Data