RAG Chat
Answers questions from files you upload, such as PDF, Word, PowerPoint or CSV. The chat model you choose writes each answer from the passages that best match the question.
Overview
RAG Chat answers questions from documents you provide. Each file is cut into chunks of a set token length, and the chunks closest to the question go to the chat model you picked along with your prompt. Models on this page include GPT-OSS 20B, Qwen3 32B, Llama 3.1 8B Instruct, Mistral Nemo, Phi-4 and the DeepSeek R1 distills.
What you can build
- Q&A over contracts and policies; the page example asks when a rental agreement starts and how much the security deposit is
- Support bots that answer from product manuals or help-center exports
- An internal helper that answers from exported mailboxes (.mbox) or Markdown wikis
Inputs
- Documents: upload several files, or list file URLs separated by commas. Accepted extensions are .pdf, .docx, .pptx, .pptm, .csv, .md, .epub, .mbox, .mp3 and .mp4. The form also takes .jpg, .jpeg and .png, but no text is extracted from images, so answers cannot draw on them. Older .ppt files are skipped, so convert them to .pptx first. The default is a sample rental agreement PDF.
- Prompt: the question itself (required), with an optional system prompt.
- Model: a public model from the list, or a private chat model from your account in its place.
- user_id and session_id: optional fields that keep each end user's conversation apart.
- Retrieval: chunk_size sets the tokens per chunk (default 256, range 10 to 8192). chunk_overlap is how many tokens neighboring chunks share (default 25). similarity_top_k is how many chunks are retrieved per question (default 5, up to 100). context_window caps the tokens the model reads at once; 0 means the model's maximum.
- Generation: max_new_tokens goes up to 4096. At 0 the reply budget is 2,048 tokens, or twice the question's length once the question runs past 2,048 tokens. Sampling starts at temperature 0.7, top_p 0.90 and top_k 50, and seed, do_sample and quantization (on by default) are also available.
Outputs
One text answer per question, written from the retrieved passages.
Recommended settings
When the facts you need sit on many different pages, raise similarity_top_k, and lower it again if answers start pulling in unrelated passages. A larger chunk_size keeps long clauses in contracts or technical manuals together. Raise chunk_overlap if answers miss text that falls on a chunk boundary.
Limitations
Only the retrieved chunks reach the model, so a question whose answer depends on several distant sections can get an incomplete reply. Check dates and amounts against the source file before relying on them.
API quick start
Run rag-chat with a single API call.
{
"prompt": "When does this agreement start? How much …",
"inputDocumentMultiple": "https://your-cdn.com/input.png",
"user_id": "...",
"session_id": "..."
}curl -X POST "https://api.wiro.ai/v1/Run/wiro/rag-chat" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "When does this agreement start? How much …",
"inputDocumentMultiple": "https://your-cdn.com/input.png",
"user_id": "...",
"session_id": "..."
}
JSONOther models you might like
Explore similar llm & chat models.



