Add a document
PDF, Markdown or text · up to 10 MB
Scanned PDFs need OCR before upload. First upload downloads the embedding model and may take a few minutes.YOUR SOURCE OF TRUTH
Knowledge base
Add approved documents. Review the extracted text. Keep answers current.
0
documents indexed
LT / EN
multilingual search
Private
no conversation storage
Import a website page
Bring an approved portal page into your knowledge base.
Review before using
Only pranesktiesa.lt pages are accepted. Import the final page URL. Menus are removed where possible; check the text preview.
Document library
CHECK THE EXPERIENCE
Test assistant
Ask questions and inspect the sources behind each result.
You are interacting with an AI assistant prototype. Do not enter identities or report details. Use the official reporting portal for submissions. Responses are general information, not legal advice.
Conversation preview
Your questions stay in this browser view and are not saved as conversation history.
SERVICE DETAILS
Configuration
Documents and retrieval run on this VPS. Generation connects through a server-side endpoint.
Current setup
RunPod connection details are configured on the server. API keys are never sent to your browser. “Configured” means settings are present; validate generation with a test question.
WORKSPACE GUIDE
Help
Turn approved documents into searchable knowledge, then test answers supported by those sources.
How the system works
- Add your sources
Upload a PDF, Markdown or text file, or import a page from pranesktiesa.lt. The system extracts its text and keeps the original content on the application server.
- Build the search index
The text is divided into overlapping excerpts. Each excerpt gets an embedding: a list of numbers representing its meaning. These vectors help find related content even when a question uses different words. They are a search index, not permanent memory inside the AI model.
- Retrieve relevant excerpts
For each question, the app combines meaning-based search with keyword matching and selects up to five excerpts from the current knowledge base.
- Generate and check the answer
The question and selected excerpts are sent to Ministral on RunPod. The model is instructed to answer only from those excerpts. The app checks that the returned source IDs match the supplied evidence before displaying the answer and source cards.
Documents are not used to train the model by this app. Uploading a source updates the searchable knowledge base. It does not retrain the model or guarantee that every answer is correct. Always review the cited text.
Getting started
- Open Knowledge base and upload an approved document, up to 10 MB. Add a clear title and an official source URL when available.
- Choose Preview to check extracted text, especially tables, dates and exceptions.
- Open Test assistant, select Lithuanian or English, and ask a specific question.
- Compare the answer with its source excerpts. Also test a question the documents cannot answer.
Keeping knowledge current
Replace updates a document and its search index together. Its previous text stops appearing in new searches.
Delete removes the document and indexed excerpts. Previously displayed answers remain in the current browser view until you clear it or reload.
Website import copies one page at a time. It does not crawl the whole site or automatically refresh changed pages. Replace or re-import outdated content yourself.
Understanding the results
AI answer: generated text with source excerpts. A source title links to the official URL if one was supplied.
No supported answer: when retrieval finds no suitable evidence, or the model cannot provide a valid supported response, the app directs you to official contacts.
Search preview: excerpts only, with no generated answer. This mode is used when model connection settings are incomplete.
Each question is independent. The model does not receive earlier chat messages, so repeat important context in follow-up questions.
Where information goes
Uploaded documents, extracted text and search vectors are stored on the VPS. Search runs there; the question and selected excerpts are sent to the configured RunPod model for generation.
The current GPU endpoint is restricted to Romania. This describes GPU placement, not a guarantee about every provider system or its retention policy.
The app does not save conversation history, and its HTTP access logging is disabled. Provider processing and retention are separate. API keys stay on the server.
Slow answers and connection issues
The GPU scales to zero when idle. A first request may take several minutes while capacity is allocated and the model starts. In our live test, the first request took about 255 seconds; subsequent requests took about 2 seconds. These are observations, not guaranteed response times.
Wait while a request is processing. If it fails, try again later. Repeated failures should be checked by the administrator.
Model configured means connection settings are present. It is not a live health check of RunPod.
Document and answer limitations
Scanned PDFs require OCR before upload. Encrypted PDFs are not supported. Text files should use UTF-8. Extraction can lose table layout or reading order, so the preview matters.
Website imports accept HTTPS pages on pranesktiesa.lt and do not follow redirects. Use the final page address shown in your browser.
Search can miss relevant information, and a valid citation does not prove that every claim is supported. Check answers against the approved sources before relying on them.
This is a private testing workspace. Do not enter identities or whistleblower report details. Submit reports through the official portal. Answers provide general information, not legal advice. This workspace is not yet a public website chat widget.