Chat with your PDFs and references, and click any citation to see the exact passage
SmartChat reads the article by section (abstract, methodology, results, references) and answers with source and page. Nine models, six specialist modes and web search when the answer is not in the file.
Free credits · No credit card · Instant access What changed recently
What this solves
SmartChat is a chat over your own material. You upload an article, a thesis chapter, a spreadsheet or a BibTeX file, or ground the conversation in your Zotero library, and ask in plain language. You choose which of the nine models answers and which specialist mode it takes, from academic advisor to data analyst. The answer arrives with a numbered source and a page number, and clicking it opens the document at that passage, highlighted. You read the evidence, not only the summary.
How it works
-
Bring the material
Drop a PDF, DOCX, TXT, CSV, XLSX or BibTeX file. SmartChat extracts the bibliographic record (title, authors, journal, year, DOI) and splits the text by page and section. You can also group sources in a project, sync a Zotero collection, or start with no document at all.
-
Choose who answers
Pick one of the nine models from OpenAI, Google and Anthropic, and a specialist mode: Advisor, Writing Assistant, Critical Reviewer, Data Analyst or Librarian. Turn on web search when the answer will not be in the file. Both choices can change from one message to the next.
-
Ask, then check the citation
Ask about the methodology and the retrieval favours the methodology section. Ask about results and it favours results, and a question about data analysis also favours passages that carry statistical terms. Each answer carries markers of source and page: click one and the PDF panel opens on that page with the quoted passage highlighted, so you confirm before citing it yourself.
Method
What happens to your file, step by step, and where the result stops being a decision made by a model.
How it works inside
The text of an uploaded file is extracted and split into 4000-character chunks with an 800-character overlap, each tagged with its page and academic section by fixed text rules, and embedded with text-embedding-3-large at 1024 dimensions. Each question scores every chunk as 0.5 times normalised keyword count plus 0.5 times cosine similarity, multiplied by 100, plus fixed boosts (+50 target section, +80 explicit table or figure reference, +30 statistical terms). The top 15 chunks, at most three per page, become the numbered [Fonte N] context.
Where AI is used, and where it is not
Retrieval is deterministic: chunk scoring, ranking, page diversity and the [Fonte N, p.X] markers added after generation follow fixed rules in code. No model decides which passage backs a claim. A language model writes the prose (OpenAI, Gemini or Anthropic, picked by the reader). Four other model calls surround that one: gpt-5.4-nano rewrites your question before it is sent, unless the question is short or a follow up; a background call reads the exchange and stores durable memories about you; and the conversation title and the follow up suggestions are written by a model as well. In web mode a model classifies the query and reranks results, and code strips [Web N] citations that match no retrieved source. The specialist mode changes nothing in the retrieval: it appends a fixed block of instructions to the system prompt.
What it accepts
- PDF, DOCX, DOC, TXT, HTML, HTM, CSV, XLSX, XLS and BIB. Both the file extension and the file's actual content type are checked. An upload request is capped at 150 MB.
- Pasted text up to 500,000 characters becomes a document. Images (JPG, PNG, WebP, GIF) up to 5 MB each, at most 5 per upload.
- A project holds at most 25 reference sources, each file up to 10 MB, from .pdf, .txt, .md, .csv, .json, .xml, .html, .htm, .doc, .docx, .xlsx, .xls.
- Zotero items reach the chat only after an explicit sync, which is a separate Pro-plan action (15 credits per sync, 2 credits per full text indexed).
- Chosen per message: one of nine models (GPT-5.5, GPT-5.4 Mini, GPT-5.4 Nano, Gemini 3.1 Pro, Gemini 3.5 Flash, Gemini 3.1 Flash Lite, Claude Opus 4.8, Claude Sonnet 5, Claude Haiku 4.5), one of six modes (Default, Advisor, Writing Assistant, Critical Reviewer, Data Analyst, Librarian) and web search on or off.
- Plans: SmartChat needs Starter, Zotero grounding needs Pro, Deep Research needs Expert.
- Billing: 15 credits to load a document, then 5, 10 or 20 credits per message depending on the model chosen in the menu. Web search adds nothing to the message. Zotero costs 15 per sync and 2 per full text, and Deep Research costs 500 per run.
What you get back
- A streamed Markdown answer rendered to HTML, with [Fonte N, p.X] badges that carry the exact sentence quoted from the chunk.
- For each citation, a reference record: the source number, the page or page range, a 150-character excerpt, and the exact position and text used to locate and highlight the passage in the PDF viewer.
- In web mode, the source list as well: title, URL, domain, relevance score, publication date and whether the answer actually cited it.
- Conversation export as PDF, DOCX or TXT, either the whole thread or only the last answer.
- Python blocks written by the model can run in a restricted sandbox (30 s, up to 3 blocks per message), returning tables and PNG charts shown directly in the chat.
What this tool does not do
- It does not judge whether the document is correct. It locates and quotes passages; whether the paper's claim holds is never checked.
- It reads only what the extractor produced. A scanned PDF with no text layer yields no chunks, and no OCR step exists.
- At most 15 chunks reach the model (20 for summary and writing requests) and the assembled context is capped at 120,000 characters, dropping to 60,000 on 10-credit models and 40,000 on 20-credit models; summary and writing requests get 25 percent more room. Text outside that window is invisible to the answer.
- In writing mode the automatic citation injection is skipped on purpose: only the citations the model itself wrote survive, so an uncited paragraph is not evidence of anything.
- If the embedding call fails, scoring falls back silently to raw keyword counting. The answer looks the same and the reader is not told.
- The question that reaches the model is not always the one you typed. A long enough question that is not a follow up is rewritten first for clarity, by a model, and the rewrite is not shown to you.
- Memories are extracted on their own, with no confirmation step. A model reads the exchange, your message and the answer, and is asked to keep only durable notes about you. Checking what it kept, and deleting it, is done in the settings panel.
- Plan
- Starter
- Cost
- 15 credits
What you get
-
Citation that opens the page
The answer cites source and page. Clicking navigates the built in PDF viewer to that page and highlights the passage the model used.
-
Six ways of answering
Default, Advisor, Writing Assistant, Critical Reviewer, Data Analyst and Librarian. Each mode loads its own set of instructions, so the same question comes back as mentoring, as criticism or as a statistical reading, depending on what you need.
-
Web search at no extra credit
When the answer is not in your file, SmartChat searches the web, ranks what it finds and cites it. A message with search on costs the same as any other message on that model.
-
Nine models, three providers
OpenAI, Google and Anthropic in the same conversation, from 5 credit models for routine questions to 20 credit ones for hard questions. If the model you chose fails, the fallback never bills above what you picked.
-
It remembers what matters
Preferences, working context and academic level are picked up as you talk and reused in later conversations. Everything it stored is listed in the settings panel, and you can edit or erase any of it.
Changelog
Every line below is a change that actually shipped, dated by the day it went out.
-
Latest
- A model fallback can no longer bill more than the model you chose was priced at.
- When only a costlier model can answer, SmartChat offers it instead of switching you silently.
- Deep Research charges before saving the answer, so a failed charge delivers nothing and costs nothing.
- The generic response error now appears in your interface language, not always in Portuguese.
Show 5 earlier updates
-
- Provider errors show a short message instead of the raw internal error on your screen.
- Conversation history and the credit balance stopped failing intermittently while you were using them.
-
- A follow-up question is now resolved against the conversation before any search or document retrieval.
- With web search on, an ordinary message no longer triggers and charges a fresh search.
- Long threads keep memory past the tenth message, with the opening question preserved.
- Attached images now reach Gemini and Claude, and your memories and project sources survive web search.
-
- Deep Research now costs 500 credits and runs on Gemini, which covers its real cost.
- During a long run the progress log shows the backend is alive instead of going silent.
- If the provider fails and the run switches to another one, the log states the reason.
-
- Your Zotero library syncs into a project and grounds answers in your own references.
- Full text of Zotero attachments is processed in the background and charged before it counts as available.
- Zotero also works in conversations outside a project, and the integration sits behind an activation switch.
- Project ownership is verified before full-text work is queued, and status no longer shows other people's counts.
-
- The answer starts sooner: preparation runs in parallel and suggestions arrive after the reply.
- Conversations keep the last ten messages, and a document contributes up to 120,000 characters.
- Web search results stream in as they are found instead of arriving all at once.
- The credits dashboard shows what a SmartChat message costs in each model tier.
Questions and answers
Which plan do I need, and what does a session cost?
SmartChat itself needs Starter. Loading a document costs 15 credits, and each message costs 5, 10 or 20 depending on the model you picked; turning web search on adds nothing. Grounding the chat in Zotero needs Pro and costs 15 credits per sync plus 2 for each full text indexed. Deep Research needs Expert and costs 500 credits per run.
What does SmartChat remember about me?
As you talk, a model picks out durable notes: a stated preference, the subject of your thesis, the level you are working at. They come back in later conversations so you do not have to repeat yourself. The full list sits in the settings panel, where each note can be edited or deleted, and the whole memory can be cleared at once.
What can it do besides answer from the uploaded file?
It searches the web and cites what it finds, grounds answers in a synced Zotero collection, reads attached images, and runs Python written by the model in a restricted sandbox, returning tables and charts in the chat. Conversations can be grouped into projects with up to 25 sources, shared by link and exported to PDF, DOCX or TXT. Deep Research, which sends an agent to read the open web for minutes and returns a report with a bibliography, also starts from this composer and has a page of its own.
Can I trust the answer?
Treat it as a pointer, not as a verdict. The model can misread. That is why every answer carries source and page, and the citation opens the passage in the document. Confirm on the page before you cite it.
Which file formats does it accept?
PDF, DOCX, DOC, TXT, HTML, CSV, XLSX, XLS and BibTeX, plus JPG, PNG, WEBP and GIF images. The page by page viewer with clickable citations works on PDFs, since it is the format that has pages.
Is my document used to train AI models?
No. We use the paid API tiers of the model providers. Under their published API policies, content sent through the API is not used for training and is retained only briefly for abuse monitoring. Our privacy policy reproduces those commitments.
Does it replace reading the article?
No. It shortens the search inside a long document and shows you where to look. Judgement about method, sample and validity stays with you, and the highlighted passage exists so you can exercise it.
How to cite this tool
Used it in your research? Here is the reference, already filled in with the version you are looking at and today's access date.
ABNT (NBR 6023)
LESSA, P. W. B. SmartChat. Versão 2026.08. [S. l.]: Xplore Dados, 2026. Disponível em: https://xploredados.com/en/tool/smartchat. Acesso em: 21 ago. 2026.
APA 7
Lessa, P. W. B. (2026). SmartChat (Version 2026.08) [Computer software]. Xplore Dados. https://xploredados.com/en/tool/smartchat
BibTeX
@software{xploredados_smartchat_2026,
author = {Lessa, Patrick Wendell Barbosa},
title = {SmartChat},
organization = {Xplore Dados},
version = {2026.08},
year = {2026},
url = {https://xploredados.com/en/tool/smartchat},
urldate = {2026-08-21}
}
Start with the free credits
Create an account and test the tools before deciding on a plan.
Create free account