Documentation
Everything you need to get started with FAQai.app.
User Guide
How to turn a PDF or other document into reviewable RAG chunks, Q&A, and evaluation data, then export them into your own system. Usage guide · API reference · Pricing
Getting Started
1. Upload a Document or Add a URL
Drag and drop or browse to upload your file (text-based PDFs, scanned PDFs (OCR, paid plans), DOCX, TXT, and Markdown (MD); size limits vary by plan — Free: 5 MB, Basic: 10 MB, Starter: 15 MB, Pro: 25 MB), or ingest content straight from a public web page or website URL / sitemap.
2. Generate Dataset
Our AI reads the full document, identifies the key topics, and generates structured training and evaluation datasets (Q&A pairs, contexts) for RAG chatbots.
3. Review & Export
Browse RAG-ready chunks, generated datasets (Q&A, variants, evaluation, adversarial), and quality insights. Export in 16 formats: JSON, CSV, Markdown, LangChain, LlamaIndex, Evaluation, Pinecone, Qdrant, pgvector, ChromaDB, Weaviate, Milvus, LanceDB, Upstash Vector, RAG Chunks, and JSONL Fine-Tuning.
Supported File Formats
| Format | MIME Type | Notes |
|---|---|---|
| application/pdf | Text-based PDFs and scanned PDFs (OCR, paid plans); file size limits vary by plan (Free: 5 MB, Basic: 10 MB, Starter: 15 MB, Pro: 25 MB) | |
| DOCX | application/vnd.openxmlformats… | Microsoft Word 2007+ format |
| TXT | text/plain | Plain text, UTF-8 encoded |
| MD | text/markdown | Markdown, UTF-8 encoded; headings improve chunk boundaries |
No file? Ingest from a URL. You can also point FAQai at a public web page, or crawl a whole same-domain site (discovered via its sitemap, or a breadth-first walk if there is none). Crawling honors robots.txt (our agent is FAQaiBot) and a per-plan page cap (Free 5, Basic 50, Starter 150, Pro 500). Only static HTML is read, so JavaScript-rendered pages may yield little content. Available in the dashboard, the REST API, the SDKs, and the MCP server.
Upload Limits
- File size: Varies by plan (Free: 5 MB, Basic: 10 MB, Starter: 15 MB, Pro: 25 MB).
- Content: Documents must contain extractable text. Scanned PDFs are supported via OCR on paid plans (Basic: 100 OCR pages/mo, Starter: 300/mo, Pro: 1,000/mo).
Uploading large files (recommended)
To upload files of any size up to your plan limit in a single request — with no infrastructure errors — use one of:
- Official SDKs:
pip install faqaiornpm install @faqai/sdk— a singleupload()call handles any file size automatically. - Upload endpoint (no SDK): POST your file to the FAQai upload endpoint, which accepts the same fields as
POST /v1/documents. Ideal for n8n, Make, and Zapier. See the large file upload guide.
Why the direct endpoints are size-limited
Direct API uploads (POST /v1/documents and POST /v1/jobs) have a 4.5 MB body size limit due to serverless infrastructure constraints. For larger files, use the SDKs or the upload endpoint above (both handle the upload transparently), or the lower-level presigned upload flow.
Plans & Quotas
| Plan | Pricing | Page Quota | RAG Playground | Features |
|---|---|---|---|---|
| Free | Free | 10 pages / month | 5 msgs/month | Core dataset generation, JSON export (max 5 MB) |
| Basic | $9/mo or $79/yr (save $29) | 1,000 pages / month | 200 msgs/month | Everything in Free + all 16 export formats + eval & adversarial datasets + API access (500 calls/mo, max 10 MB) |
| Starter | $19/mo or $190/yr (save $38) | 2,000 pages / month | 500 msgs/month | Everything in Basic + bulk upload + 1,000 API calls/mo (max 15 MB) |
| Pro | $39/mo or $390/yr (save $78) | 6,000 pages / month | 1,500 msgs/month | Everything in Starter + webhooks + 3,000 API calls/mo (max 25 MB) |
Free plan quotas reset on the 1st of each calendar month. Paid plan quotas follow your Stripe billing cycle.
Yearly plans receive the same monthly page quota as monthly plans (quotas reset each month, not granted upfront). Yearly billing saves approximately 2 months of cost.
Website Crawl Pages per Crawl
When you ingest from a website URL, a single-page fetch counts as one page, and a same-domain crawl fetches up to this many pages per crawl. Extracted content counts against your monthly page quota above (static HTML only, robots.txt-aware).
Free
5 pages
Basic
50 pages
Starter
150 pages
Pro
500 pages
What you receive
FAQai prepares datasets from files you upload or from a public URL you are allowed to ingest. It does not host your production chatbot or your vector database. You export the dataset and import it into your own stack.
Exports include embedding_text, the text to embed. They do not include precomputed embedding vectors. The LlamaIndex chunk export sets embedding to null. You generate vectors with your own model. The usage guide shows that step. The Upstash Vector format puts the same text in data so Upstash can embed it during import. That is the server-side exception.
There are 16 export formats and 8 vector-database targets: Pinecone, Qdrant, pgvector, Chroma, Weaviate, Milvus, LanceDB, and Upstash. Chunking presets are Balanced (~400 tokens, the default), Fine-grained (~200), and Large context (~800). Custom target size (128–2048 tokens), overlap (0–25%), and strategy (heading-aware, fixed size, or paragraph) are paid-plan controls. The Free plan uses Balanced.
Coverage and groundedness are review signals on the dataset you generated. The RAG Playground answers from that document so you can inspect sources. Neither is proof that a production system is accurate or free of hallucinations.
The remote MCP endpoint is https://faqai.app/api/mcp with OAuth and 17 tools. The local stdio package has 18 tools because only stdio can upload_document. API access, including remote MCP, requires Basic, Starter, or Pro. Completion webhooks require Pro. Tools such as n8n can call the API or receive webhooks. That is not a built-in FAQai connector. Plan limits are on the pricing page.
Export Options
FAQai supports 16 export formats. JSON is on every plan. The other formats require Basic, Starter, or Pro.
| Format | Description | Plans |
|---|---|---|
| JSON | Standard format with all metadata | All plans |
| CSV | Spreadsheet-compatible for bulk editing | Basic, Starter, Pro |
| Markdown | Human-readable format for documentation | Basic, Starter, Pro |
| LangChain JSON | Compatible with LangChain document loaders | Basic, Starter, Pro |
| LlamaIndex JSON | Compatible with LlamaIndex data structures | Basic, Starter, Pro |
| Evaluation | RAGAS / DeepEval compatible evaluation format | Basic, Starter, Pro |
| Pinecone | Ready for Pinecone vector database upsert | Basic, Starter, Pro |
| Qdrant | Ready for Qdrant vector database | Basic, Starter, Pro |
| pgvector | SQL INSERT statements for PostgreSQL pgvector | Basic, Starter, Pro |
| ChromaDB | Parallel arrays format for ChromaDB collections | Basic, Starter, Pro |
| Weaviate | Objects format for Weaviate batch import | Basic, Starter, Pro |
| Milvus | Rows format for Milvus collection insert | Basic, Starter, Pro |
| LanceDB | JSON array format for LanceDB tables | Basic, Starter, Pro |
| Upstash Vector | Auto-embed format for Upstash Vector index | Basic, Starter, Pro |
| JSONL Fine-Tuning | OpenAI / Mistral / Together AI fine-tuning format | Basic, Starter, Pro |
You can also copy individual items to clipboard from the dataset browser. JSON export includes all dataset types; other formats export the selected dataset type.
RAG Chunks Export
The RAG Chunks tab has its own dedicated export dropdown supporting all 16 formats. Chunk exports include enriched metadata (keywords, entities, topic, section), quality signals, and retrieval hints optimised for direct ingestion into vector databases and RAG pipelines.
JSON Export Diagnostics
JSON exports include two diagnostic sections: failure_analysis (RAG quality issues such as ambiguous queries, chunking gaps, and semantic overlap) and internal_errors (pipeline-level diagnostics like variant count or parsing issues). These help identify and fix weaknesses in your RAG system.
Dataset Types
FAQai generates five dataset outputs: RAG Chunks, Canonical Q&A, Query Variants, Evaluation, and Adversarial. Quality Insights is a separate quality analysis, not a sixth dataset type.
RAG Chunks
Structured, enriched document chunks with metadata (keywords, entities, topic, section), quality signals (completeness, overlap risk, fragmentation risk), and retrieval hints (search type, priority, use-case tags). Optimised for feeding into RAG pipelines.
Canonical Q&A
Core question-answer pairs grounded in document content. Each includes context, confidence score, and question type classification.
Query Variants
10 alternative phrasings per canonical question across 8 variant types (short, vague, broken, typo, incorrect assumption, partial, conversational, multi-intent) to improve retrieval robustness.
Evaluation Dataset
Question-context-expected_answer triples for retrieval and generation evaluation. Compatible with RAGAS and similar frameworks. Available on paid plans only.
Adversarial Dataset
Misleading and edge-case questions to test RAG system robustness against hallucination and out-of-scope queries. Available on paid plans only.
Quality Insights
Not a dataset type. A separate analysis of coverage and weaknesses, with risk levels and recommendations.
Pages per Month
Usage is measured by pages processed per document. Paid plans can purchase overage page packs that also include bonus RAG Playground messages.
Free
10 pages
Basic
1,000 pages
Starter
2,000 pages
Pro
6,000 pages
Overage Packs (Paid Plans)
Purchase additional capacity when you need it. Each pack adds extra pages and bonus RAG Playground messages for the current billing cycle.
| Pack | Pages | Price | Playground Bonus |
|---|---|---|---|
| Small | 1,500 | $9 | +500 messages |
| Medium | 5,000 | $29 | +1,500 messages |
| Large | 10,000 | $49 | +3,000 messages |
Dataset Items per Document (Per-Type Caps)
Each dataset type has its own guaranteed allocation so no single type can starve the others.
| Dataset Type | Free | Basic | Starter | Pro |
|---|---|---|---|---|
| Canonical Q&A | 80 | 200 | 400 | 600 |
| Query Variants | 200 | 500 | 1,000 | 1,500 |
| Evaluation | — | 150 | 300 | 450 |
| Adversarial | — | 150 | 300 | 450 |
Browsing Datasets
Once a document is processed, you can browse the result on the document detail page. The five dataset tabs are RAG Chunks, Canonical Q&A, Query Variants, Evaluation, and Adversarial. Quality Insights, RAG Config, and RAG Playground are separate tabs.
Document Tabs
Switch between tabs to view items from each dataset type. Each tab shows its item count. All dataset tabs share the same search bar.
RAG Config provides an AI-generated system prompt, embedding model recommendations, and ready-to-use code snippets. RAG Playground lets you chat with your document to test retrieval quality interactively - see the RAG Playground section below.
Confidence Scores (Canonical Q&A)
Each Canonical Q&A item displays a confidence score indicating how directly the answer is supported by the document. Hover over the badge for a detailed explanation.
| Level | Meaning |
|---|---|
| High | Answer is directly stated in the document |
| Medium | Answer is supported but requires some inference |
| Low | Answer requires significant inference from the document |
Difficulty Badges (Evaluation & Adversarial)
Evaluation and Adversarial dataset items display a difficulty badge. Hover over the badge for a detailed explanation.
| Level | Meaning |
|---|---|
| Easy | Straightforward question answerable from a single passage |
| Medium | Requires combining information from multiple parts of the document |
| Hard | Requires deep reasoning, inference, or handling ambiguous information |
Pagination
Dataset items are paginated at 25 items per page across all tabs. Use the page number buttons or Previous / Next to navigate. The search bar filters items across question and answer text within the active tab.
RAG Playground
The RAG Playground lets you chat with any processed document to test retrieval quality interactively. Ask questions and see how the AI responds using your document's content - before deploying to production.
How to access
Open any processed document and switch to the RAG Playground tab. Type a question and press Enter to chat with your document.
Monthly message limits
Each plan includes a monthly RAG Playground message allowance. There is no per-document limit - you can spread your messages across any documents.
Messages per Month
Free
5 msgs
Basic
200 msgs
Starter
500 msgs
Pro
1,500 msgs
Overage Bonus
Paid plans can purchase overage page packs that also include bonus RAG Playground messages: Small (+500 msgs), Medium (+1,500 msgs), Large (+3,000 msgs). Bonus messages are valid for the current billing cycle.
Your remaining messages are shown in the Playground tab header and on your Dashboard. Message limits reset at the start of each billing cycle.
OCR for Scanned PDFs
FAQai uses a hybrid OCR approach powered by a vision LLM model to extract text from scanned or image-based PDFs. This feature is available on paid plans only.
How It Works
- Native extraction first: FAQai attempts standard text extraction from each PDF page.
- Automatic OCR fallback: Pages without extractable text are automatically sent to the vision LLM for OCR.
- Hybrid output: The final document combines natively extracted text and OCR-processed pages seamlessly.
OCR Page Caps per Month
Free
0 pages
Basic
100 pages
Starter
300 pages
Pro
1,000 pages
The API response includes extraction_method (native / hybrid / full_ocr) and ocr_page_count fields so you can track OCR usage per document.
Cancel Processing
You can cancel a document that is currently being processed. Cancellation stops dataset generation and refunds the page credits back to your monthly quota.
From the document list
Click the three-dot menu on a processing document and select "Cancel Processing".
From the document detail page
Click the amber "Cancel Processing" button next to the Delete button while the document is processing.
Via API
Send POST /v1/documents/{id}/cancel to cancel programmatically. See the API Reference section.
Cancelled documents appear with a “Cancelled” status badge. Any partially generated datasets are marked as failed and excluded from exports.
Security & Privacy
Your uploaded documents are processed securely using TLS encryption in transit and AES-256 encryption at rest. Document text is sent to AI models for dataset generation and is not used for model training.
All data is protected by row-level security policies - your documents and datasets are only accessible by your account.
For more details, see our Privacy Policy, Security Page, and GDPR Compliance.