Why does AI chat usually give you wrong answers about documents?
Most AI document tools summarize your file first, then answer questions from that summary. This is fast and cheap, but it means the AI is answering from a lossy compression of your document, not the document itself. Details, nuance, and anything the summarizer deemed unimportant are gone.
The other common failure is fabricated citations. When a model does not know which page something came from, it guesses. Students trust the citation, look up the page, and find nothing there. This is a subtle but serious reliability problem.
Recitare's Ask is built differently. The full text of your document is sent to the model with every query. Recitare is instructed to cite only page boundaries that actually exist in the parsed text, and to acknowledge uncertainty when it does not know the exact location of a claim.
How does Ask work?
The moment you open a document in Recitare, the full parsed text is available for chat. Here is the end-to-end flow:
Open a document
Upload a PDF, DOCX, TXT, PPTX, image, or paste a URL. Recitare parses the full text and inserts page markers (--- Page 1 ---, --- Page 2 ---) so the model knows exactly where each page begins.
Open the chat panel
Choose Ask Max above the document, or Ask in the bottom bar on a phone. The panel slides in on the right. Your document is already loaded, no extra upload step.
Ask anything
Type any question: "What methodology did they use?", "Summarize the limitations section", "Find every mention of dopamine", "What is the author's main argument?" Recitare reads the entire document to answer.
Get a cited answer
Recitare responds with the answer and [Page X] references for every claim. Click a citation to jump directly to that page in your document.
Keep asking
Follow-up questions carry full conversation context. Recitare remembers everything you've discussed in the current session, so you can drill down, push back, or ask for clarification without repeating yourself.
Does Recitare actually read the whole document?
Yes, and this is the detail that matters most. Ask is powered by Gemini, and the whole document you opened is included in the prompt, not an excerpt or a summary of it. In practice that covers what you would upload: undergraduate papers, theses, technical manuals, course readers.
For documents longer than 20,000 characters (around 5,000 words), Recitare uses context caching. Your document is cached with Google's API for 30 minutes. Subsequent questions in the same session skip the re-processing cost and come back faster. This is entirely invisible to you. The chat just gets snappier as the session goes on.
How accurate are the citations?
Recitare's PDF parser inserts explicit page markers into the text as it processes each document: --- Page 12 ---. These markers are included in the text sent to Gemini, so the model knows the exact page boundary of every sentence. When Recitare cites [Page 12], it is referencing a real boundary in your document, not estimating.
Recitare is also instructed to express uncertainty honestly. If the relevant content spans multiple pages or Recitare cannot pinpoint the exact page, it will say so rather than fabricate a confident but wrong citation.
- Citations are grounded in explicit page markers from the parser, not inferred
- Recitare acknowledges uncertainty instead of guessing
- Every answer links to verifiable text in your document
- No hallucination of facts not present in the source
What is Ask While Listening?
One of the most common friction points in active reading is the moment you want to ask a question but don't want to stop your listening session to type. Ask While Listening solves this.
A floating "raise your hand" button appears in the bottom corner of the screen whenever TTS playback is active. Tap it and:
Playback pauses at once
The TTS stops at the exact sentence it was reading. Your position is saved.
Chat panel opens
Type your question. Recitare answers with the full document context, not just the section you were listening to.
Dismiss and resume
Close the chat panel and TTS picks up from the exact sentence where it paused. No scrubbing, no losing your place.
Ask While Listening is available on Scholar ($9/mo) and Researcher ($18/mo) plans, where TTS playback is API-powered and supports precise sentence-level pause and resume.
What is different about Ask?
Document chat is not new. What matters is where the answer comes from and whether you can check it. Ask works from the whole document you opened, answers from that text rather than from general knowledge, and cites the page each answer rests on. The citations come from the parser that read your file, so they point to real pages. And it sits inside the reader: you ask while listening, and playback resumes where it paused.
What is happening under the hood?
For those who want the technical picture:
Document parsing
PDFs are parsed client-side with pdfjs-dist. DOCX, TXT, and PPTX files are parsed server-side. All formats produce plain text with --- Page N --- markers. Images are OCR'd with Tesseract.js.
Gemini with context caching
The full document text + conversation history is sent to Gemini on every request. Documents over 20K characters are cached for 30 minutes, giving a 75% input token cost reduction on repeat queries.
Citation grounding
The system prompt instructs the model to cite [Page X] using the --- Page N --- markers in the document text. The model is trained to prefer explicit evidence over inference and to express uncertainty when it cannot locate the exact page.
Conversation history
Each API call to /api/chat includes the full message history from the current session. The model receives: system prompt + document text + prior Q&A pairs + new user question. This allows genuine multi-turn conversations.
Who is Ask for?
Students reading dense papers
Ask about a concept before you read it, then ask follow-up questions as you go, so jargon stops being a wall.
Researchers doing literature reviews
Upload a paper and ask "what does this add that the 2022 Ioannidis study doesn't cover?". Triage a stack of papers before deciding which to read in full.
Auditory learners using TTS
Listen to a lecture or paper while commuting, then ask clarifying questions mid-listen without losing your place.
Professionals reading reports
Ask "what does this say about Q3 performance?" or "find the risk factors section" without manually scanning a 60-page PDF.
Frequently asked questions
Does Recitare actually read the whole document?
Yes. Your whole document is included in the prompt, not a summarised excerpt, so a question about a footnote on page 84 is answered from page 84.
How accurate are the page citations?
Very accurate. Recitare is instructed to cite [Page X] using the page markers that Recitare's PDF parser inserts into the text. When Recitare says "see Page 12", it is referencing the actual page boundary from your document. If Recitare is not confident about which page something appeared on, it will say so rather than hallucinate a citation.
Do follow-up questions remember the conversation?
Yes. Every message in the chat panel is sent with the full conversation history, so Recitare remembers what you asked before. You can say "tell me more about that" or "how does that relate to what you said earlier" and Recitare will follow the thread. The conversation builds on itself like talking to a tutor who has been in the room the whole time.
What is Ask While Listening?
Ask While Listening is a floating button that appears during TTS playback. Tap it and playback pauses so you can type a question. Recitare answers in the chat panel, and when you dismiss the answer, playback resumes from exactly where it stopped. It means you do not have to choose between listening and asking.
How many questions can I ask per day?
Free users get 5 chat messages per day. Paid subscribers (Scholar at $9/month and Researcher at $18/month) can ask as much as they need within a fair-use cap of 1,000 messages per month.
Does it work with all document types?
Yes. Document chat works with every format Recitare supports: PDFs, DOCX files, TXT files, PowerPoint slides (PPTX), images processed via OCR, and web content added via URL. The parsed text from all formats flows through the same chat pipeline.
What is context caching and why does it matter?
For documents over 20,000 characters (~5,000 words), Recitare caches the document context with Google's API for 30 minutes. This means your second, third, and subsequent questions in a session cost 75% less to process than the first. It also makes follow-up responses faster because the model does not re-read the full document every time.
Can Recitare answer questions about charts or figures in my PDF?
Recitare can discuss chart data and figure descriptions when that information is represented as text in your document (captions, data tables, surrounding prose). For visual figures that exist only as images in the PDF, use the Snap-to-Explain Snap-to-Explain: draw a rectangle over the figure and Recitare will analyze it at 300 DPI via vision AI.