Transform scattered documents into secure, searchable knowledge. Ask questions, discover insights and get AI-powered answers backed by verifiable sources.
- Workspace-scoped retrieval
- Hybrid vector + keyword search
- Answers with verifiable sources
Documents
Drop files to add them to this workspace
Drop PDF, DOCX, CSV or XLSX
Up to 50 MB per file
- 0%
FY24-annual-report.pdf
- 0%
supplier-contracts.docx
- 0%
q4-revenue-model.xlsx
The problem
Your documents already hold the answer. Finding it is the hard part.
Most teams do not have a knowledge problem. They have a retrieval problem.
Today
Knowledge is scattered
Contracts sit in one drive, policies in another, models in a third. Nobody knows which version is current.
Search returns files, not answers
Keyword search finds a 200-page PDF. It does not tell you what page 147 actually says.
General AI tools cannot see your documents
A chatbot trained on the public internet has never read your supplier agreement — so it guesses.
With Askivyo
One indexed knowledge base
Every document is parsed, chunked, embedded and indexed with its structure preserved — pages, sections, sheets and rows.
Answers with the passage attached
Each claim carries a citation that opens the exact page, section or spreadsheet range it came from.
Grounded, or it says so
Retrieval runs before generation. When the evidence is thin, Askivyo tells you rather than inventing an answer.
How it works
From raw file to cited answer in four steps
Each step is observable. You can watch a document move through the pipeline and see exactly where it is.
Upload
Drag in PDF, DOCX, CSV or XLSX files. Askivyo verifies the real file type, scans it and stores the original in private object storage.
Process
A background pipeline extracts text, runs OCR on scanned pages, preserves structure, chunks by meaning and generates embeddings.
Retrieve
Questions run vector and keyword search in parallel, fuse the results and rerank them — all inside your authorised document scope.
Answer
The model writes from the retrieved passages only, and every claim links back to the page it came from.
Document intelligence
Built for documents, not for chat transcripts
The quality of a RAG answer is decided long before the model runs. Most of Askivyo's engineering sits in the ingestion pipeline.
Structure-aware extraction
PyMuPDF for PDFs, python-docx for Word, openpyxl and pandas for spreadsheets. Headings, pages, sheets and row ranges survive extraction.
OCR for scanned pages
Pages with too little extractable text are routed through a pluggable OCR adapter, and the document records that OCR was applied.
Semantic chunking
Content is split on structural boundaries with token-bounded overlap, so a chunk is a coherent passage rather than an arbitrary slice.
Versioned indexing
Parser, chunker and embedding versions are recorded per document version, so an answer given last month stays reproducible.
Streaming conversations
Answers stream token by token over server-sent events, with sources delivered before the first word so you can read along.
Deletion that actually deletes
Removing a document purges its chunks and embeddings and deletes the stored object, so it cannot resurface in a later answer.
Hybrid retrieval
Two searches are better than one
Vector search understands meaning but misses exact identifiers. Keyword search nails an invoice number but has no idea that “churn” and “attrition” are the same idea. Askivyo runs both and fuses the rankings.
- pgvector similarity
- Cosine distance over embeddings stored in PostgreSQL, with an IVFFlat index and the authorised document set applied inside the scan.
- PostgreSQL full-text search
- A generated
tsvectorcolumn ranked withts_rank_cd, catching literal terms embeddings smooth over. - Reciprocal rank fusion, then reranking
- Both rankings merge with RRF, and a reranking model scores the survivors before anything reaches the model's context window.
Vector
- Risk factors · p.310.87
- Financial review · p.120.82
- Outlook · p.440.71
Keyword
- Financial review · p.120.91
- Pricing schedule · §4.20.68
- Glossary · p.510.44
Fused + reranked
- Financial review · p.120.98
- Pricing schedule · §4.20.74
- Risk factors · p.310.69
Source-grounded answers
An answer you cannot verify is not an answer
Askivyo is built so that trusting the output does not require trusting the model.
Every citation resolves
A citation stores a real chunk ID and document version. Clicking it opens that exact passage — there is no way to render a citation that does not exist.
Quotes are verified substrings
The quoted passage is checked against the stored chunk text before the citation is saved. A quote the model paraphrased is rejected.
Abstention over invention
If reranked relevance falls below a configured threshold, the assistant says the documents do not contain enough evidence.
Retrieved text is untrusted
Document content is wrapped in delimited evidence blocks with instruction-like patterns neutralised, so a malicious PDF cannot redirect the model.
Collaboration
Shared knowledge, separated by design
Teams need the same knowledge base without seeing each other's confidential material. Askivyo separates the billing boundary from the retrieval boundary so both are possible.
- Organisations hold billing, members and audit history.
- Workspaces scope retrieval — an answer never crosses a workspace boundary.
- Four roles: owner, admin, editor and viewer, enforced on the server.
- Restricted documents can be narrowed below workspace level with explicit grants.
- Invitations are single-use, expiring tokens tied to a role.
- Every permission change is written to an append-only audit log.
A question asked in Research cannot retrieve a chunk from Project Ash, even for an organisation owner, unless they are a member of that workspace.
Security
Access control is the feature
For a document AI product, isolation is not a checkbox at the bottom of the page — it is the thing that makes the product usable at all.
Read the security overviewAuthorisation before retrieval
The authorised document set is computed first and applied as a SQL predicate, not filtered out afterwards.
Private object storage
Originals live in a private bucket. Downloads use short-lived signed URLs that are re-authorised on every request.
Row-level security
PostgreSQL RLS policies enforce tenant isolation independently of application query construction.
Sessions and rate limits
Rotating refresh-token families in httpOnly cookies, with per-route rate limits on auth, upload and chat.
Use cases
Different documents, same problem
Askivyo does not ship industry-specific models. It ships an ingestion pipeline and a retrieval layer that work on whatever your team already writes.
Diligence across hundreds of filings
Pull covenant terms, margin movements and risk disclosures out of annual reports, memos and models — with the page reference attached for the investment committee.
Questions teams ask
- What changed in the covenant package between the FY23 and FY24 facility?
- Summarise every mention of customer concentration risk.
- Which assumptions differ between the base and downside model?
Pricing
Plans that scale with your document volume
Limits are enforced by the backend, not by the interface. Every plan includes the full retrieval pipeline — the difference is capacity and access control.
Free trial
Evaluate the full pipeline on your own documents.
14 days, no payment details
Start free trial- All ingestion formats: PDF, DOCX, CSV, XLSX
- Hybrid search and grounded AI answers
- One workspace, up to 3 members
- No payment details required
- Documents
- 25
- Storage
- 500 MB
- AI messages / month
- 200
- Seats
- 3
Team
Most capacity per seatFor a single department running a shared knowledge base.
- Unlimited workspaces
- Collections and document permissions
- Usage analytics and cost tracking
- Email support
- Documents
- 1,000
- Storage
- 25,000 MB
- AI messages / month
- 5,000
- Seats
- 15
Business
For organisations with several teams and stricter access rules.
- Everything in Team
- Restricted documents with per-user grants
- Full audit log export
- Configurable data retention
- Google Workspace sign-in
- Documents
- 10,000
- Storage
- 250,000 MB
- AI messages / month
- 50,000
- Seats
- 100
Enterprise
For deployments with bespoke security, scale or hosting requirements.
- Everything in Business
- Self-hosted or dedicated deployment
- Custom model and embedding providers
- Bring-your-own object storage and KMS keys
- Named technical contact
- Documents
- Custom
- Storage
- Custom
- AI messages / month
- Custom
- Seats
- Custom
About these figures
FAQ
Questions worth asking before you trust a document AI
Generation only ever sees passages that were retrieved from your documents, wrapped in labelled evidence blocks. Each claim must cite a chunk, and the quoted passage is validated as a real substring of that chunk before the citation is stored. When reranked relevance falls below the configured threshold, the assistant abstains and says the documents do not contain enough evidence.
Put your documents to work
Start a free trial, upload a handful of documents and ask the question you have been asking colleagues all week.
No payment details required for the trial.