Ingestion and parsing
Upload directly or connect object storage. Documents are parsed with layout preserved, so tables stay tables and headings stay attached to their sections.
NXAARA Platform service
Answers from your documents, with the source attached
Ingestion, OCR, chunking, embeddings, vector search and reranking, managed end to end — producing answers that cite the document and page they came from, so a reader can check them.
Retrieval is a pipeline, and every stage can quietly ruin it
Teams building their first RAG system usually discover the same thing in the same order. The demo works. Then somebody uploads a scanned PDF and it returns nothing. Then somebody asks a question phrased differently from the document and it returns the wrong section. Then somebody asks a question the corpus does not answer and it confidently invents something.
None of these are retrieval problems in the abstract. They are ingestion, chunking, reranking and grounding problems, and each one is a separate piece of engineering that has to be built and tuned.
Knowledge Cloud runs that pipeline as a managed service. OCR for scanned documents including Arabic. Layout-aware chunking that does not split a table in half. Embeddings and vector storage. A reranking pass that reorders retrieved passages by actual relevance rather than raw vector distance. And grounding, so that when the corpus does not contain an answer the response says so instead of improvising.
Every answer carries citations to the document and page. Not as a nicety — as the mechanism by which a person can verify the answer without trusting the model.
- Broad ingestion — PDF, Office documents, HTML, plain text, images, with connectors to object storage
- OCR including Arabic — scanned and photographed documents, including mixed-script pages
- Layout-aware chunking — respects headings, tables and section boundaries instead of cutting at a character count
- Reranking — a second pass that reorders results by relevance, which is usually the single largest quality gain
- Citations — every answer points at the document and page it came from
- Permission-aware retrieval — a user only retrieves from documents they are allowed to see
Capabilities
The pipeline, stage by stage
Each of these is a place where retrieval quality is won or lost, which is why they are managed rather than left to you.
OCR
Scanned and photographed documents, English and Arabic, including pages that mix both scripts and pages that arrive slightly rotated.
Chunking and embedding
Chunk boundaries follow document structure. Embeddings are generated with a model appropriate to the language and domain of the corpus.
Hybrid search
Vector similarity combined with keyword matching, because exact identifiers, part numbers and legal references are exactly what pure vector search misses.
Reranking
A cross-encoder pass reorders candidates by real relevance. This is unglamorous and it is typically the biggest single improvement in answer quality.
Grounded generation
The answer is generated only from retrieved passages, with citations attached. When the corpus has no answer, the response says so.
How it works
Standing up a knowledge base
Retrieval quality compounds through this sequence, so each step is worth getting right before moving on.
Create a knowledge base
A scoped collection with its own access rules, embedding configuration and retention policy.
Ingest documents
Upload or connect a source. Parsing, OCR, chunking and embedding run as one job with a status you can watch.
Test retrieval
Query it directly, before any generation. Look at which passages come back and in what order — most quality problems are visible here.
Tune and rerank
Adjust chunk size, hybrid weighting and reranking depth against a set of real questions from real users.
Connect to an endpoint
Attach the knowledge base to an inference endpoint so answers are grounded and cited in production.
Why citations are a requirement and not a feature
A generated answer with no source is a claim. A generated answer with a citation is a claim a reader can check in ten seconds. In any setting where being wrong has a consequence — policy, medical, legal, financial, engineering — that difference is the whole basis on which the system can be trusted at all.
It also changes how failures behave. An uncited system that retrieves the wrong section produces a confident wrong answer that nobody catches. A cited system that retrieves the wrong section produces an answer whose citation obviously does not support it, and a user notices immediately.
This is why grounding is enforced rather than encouraged. If the retrieved passages do not support an answer, the correct output is to say the corpus does not cover it. That is a useful response. An invented one is not.
| Document formats | PDF, DOCX, PPTX, XLSX, HTML, TXT, Markdown, images |
|---|---|
| OCR languages | English and Arabic, including mixed-script pages |
| Chunking | Layout-aware, with configurable size and overlap |
| Search | Hybrid vector plus keyword, with configurable weighting |
| Reranking | Cross-encoder pass with configurable depth |
| Citations | Document, page and passage returned with every answer |
| Access control | Document-level permissions honoured at retrieval time |
| Updates | Incremental re-ingestion; changed documents are re-embedded without rebuilding the base |
Where it is used
Where grounded retrieval matters most
Policy and procedure assistants
Staff asking what the rule is, getting the answer and the clause it came from, so they can read it themselves when it matters.
Technical and maintenance documentation
Engineers searching manuals and service bulletins where the exact part number and the exact revision are the whole question.
Contract and tender review
Finding the clause, the obligation and the deadline across hundreds of documents, with a citation an analyst can verify.
FAQ
Common questions
How is this different from just uploading files to a chatbot?
A file upload puts a document into a context window. A knowledge base parses, OCRs, chunks by layout, embeds, indexes, retrieves with hybrid search, reranks, and cites. Those stages are why it still works at ten thousand documents instead of ten.
Does it handle scanned Arabic documents?
Yes, including pages that mix Arabic and English. Scanned quality still matters — a clean scan reads far better than a photograph taken at an angle — but Arabic OCR is a first-class part of the pipeline rather than an afterthought.
What happens when the answer is not in the documents?
The system says the corpus does not cover it. Grounding is enforced, so the generation step only works from retrieved passages. An honest gap is a far better outcome than a fluent invention.
Can different users see different documents?
Yes. Permissions are applied at retrieval time, not filtered afterwards, so a user's query only ever searches documents they are entitled to see. A restricted document cannot leak through a summary.
How do I know retrieval quality is good enough?
Test it with real questions from real users before connecting generation. Look at which passages come back and in what order. If the right passage is not in the top few results, no amount of prompt engineering downstream will fix it.
Ingest a hundred documents and query them
Retrieval quality on your own corpus is the only test that means anything. Start with a real subset rather than a sample set.