Knowledge Base
Make it answer from your documents.
Connect sixteen kinds of source, choose how documents are indexed, and tailor retrieval with hybrid search, HyDE, cross-encoder reranking, Self-RAG, CRAG, or Agentic RAG. Test the results in the playground before bringing them into customer conversations.
Answers with something real behind them.
Bring documents and supported sources into the knowledge pipeline.
01 / 0316
source types
5
chunking strategies
4
embedding providers
9
live retrieval strategies
How an answer gets grounded
Retrieval is where a bot stops making things up.
A knowledge base is not a folder you upload and hope about. It is a pipeline — chunk, embed, index, search, filter — and every decision in it changes what your customer is told.
Connect the source
Upload files, or connect cloud storage, a SaaS workspace, a database, an API, or a crawl — sixteen source types in total.
Chunk and embed
The document is split by the strategy you chose and embedded with your model, then indexed into a vector store scoped to your workspace.
Retrieve on the turn
A real question embeds and searches. Dense by default; add keyword fusion, a query rewrite, or a reranker when the answer quality demands it.
Drop the weak evidence
Chunks below the relevance floor are never injected. An off-topic question gets no context rather than the six least-irrelevant paragraphs.
Ground the answer
What survives is handed to the model as evidence, and the whole retrieval shows up in the trace for that turn.
Sixteen sources
Your knowledge already lives somewhere.
Point Rylvo at it instead of migrating it. Each source keeps its own sync schedule and reports its own status, so a stale knowledge base is something you can see rather than something you find out about from a customer.
Upload
No auth- File upload
Cloud storage
Token- Amazon S3
- Google Cloud Storage
- Azure Blob
SaaS
OAuth or token- Notion
- Confluence
- Google Drive
- SharePoint
- Zendesk
- Intercom
- Slack
Database
Connection string- PostgreSQL
- MySQL
- MongoDB
API & web
Token or none- REST API
- Web crawl
Sync on a schedule, or only when you say so
Realtime, hourly, daily, weekly, or manual — per source. Each one reports whether it is connected, syncing, synced, paused, or in error.
Chunking and embedding
Most bad retrieval is really bad chunking.
How a document is cut decides what can ever be found in it. Rylvo makes that an explicit choice rather than a hidden default, and tells you plainly that changing it means a reindex.
Five chunking strategies
Chosen per connection, applied at ingest.
Recursive
Splits on paragraph, then line, then word boundaries, and packs to size. The general-purpose default.
Sentence
Packs whole sentences up to the chunk size. Good for prose, articles, and FAQs.
Fixed-size
Hard token windows with overlap, ignoring text boundaries. Predictable — good for logs and transcripts.
Structure-aware
Splits on Markdown headings first, then by size inside each section.
Whole document
One chunk per document, for short self-contained docs where full context matters most.
Your embedding model
OpenAI, Voyage, Cohere, or Jina with your own key, or the indexing included with your plan. The query path uses the same embedding model that indexed the source for consistent retrieval.
Chunk size and overlap
Set the target tokens and the overlap per connection, and choose between flat chunks or a parent-child structure.
Retrieval strategies
Choose a retrieval path. Make every answer count.
Start with proven search, then add advanced methods where your content and evaluation call for them.
Start with a strong retrieval foundation
Dense search is the default. Add keyword fusion or an indexing profile when your content calls for a more tailored search path.
Hybrid search
Combine dense vectors and keyword search with reciprocal rank fusion to improve recall for exact terms such as product codes.
Classic RAG
Dense similarity search with a relevance floor: a straightforward foundation for document Q&A.
Parent-child
Match on small precise chunks, then hand the model the parent for context. Resolved at ingest.
Contextual retrieval
Prefix each chunk with document context before embedding to preserve meaning across isolated passages.
Parent-child and contextual retrieval are applied at ingest, so retrieving over them costs nothing extra at query time. Changing either one means a reindex.
Prove it before you ship
See the chunks the model will actually be handed.
Retrieval quality is not a feeling. Run the query, read what came back, and check the score of the chunk that convinced your bot to say what it said.
KNOWLEDGE BASE / PLAYGROUND PREVIEW
ILLUSTRATIVE · HYBRID + RERANKQuery
can I return a jacket I bought three weeks ago?
returns-policy.md
Items may be returned within 30 days of delivery in original condition…
returns-policy.md
Refunds are issued to the original payment method within 5 business days…
shipping-faq.md
Delivery estimates are calculated at checkout and exclude public holidays…
warranty.md
below floor · droppedExtended cover is sold separately and is not part of the standard offer…
3 CHUNKS SENT TO THE MODEL
The playground
Run a real query against a connection and see every chunk it returned, with its score, its source document, and the latency it cost.
The relevance floor
Chunks below the similarity floor are dropped rather than injected, and a query too short to embed meaningfully skips retrieval entirely instead of matching noise.
The trace
On a production turn, the retrieval is part of the observation tree — what was searched, what came back, and what the model was given.
See the evidence behind each answer
Run a question in Test, review the retrieved passages, and use the trace to see what the model received on a live turn.
Access and plans
One source is enough to find out if this works.
Every plan can ground a bot in real knowledge. What scales is how many sources you connect and whether you can build your own retrieval pipelines.
Free
1 source
No custom blueprints
Lite
15 sources
5 blueprints
Pro
50 sources
25 blueprints
Team
50 sources
25 blueprints
Enterprise
Unlimited
Unlimited
Bring your own embedding key, or use the indexing included with your plan
Index with your own OpenAI, Voyage, Cohere, or Jina key and the cost is yours to see, on your own account. Or use the indexing included with your plan, up to its monthly allowance. Reranking uses your own Cohere, Voyage, or Jina key.
FAQ
What runs, what reindexes, and what falls back
The details that decide whether your bot answers from your knowledge or from its imagination.
What actually happens to my document?
It is split by the chunking strategy you picked, embedded with your chosen model, and indexed into a vector store scoped to your workspace. Changing the chunking or the embedding model means a reindex, because the old vectors were built under different rules.
Which embedding models can I use?
OpenAI, Voyage, Cohere, and Jina embedding models, with your own provider key — or use the indexing included with your plan, up to its monthly allowance, on the default OpenAI model. The query path embeds with the same model the source was indexed with — a mismatch would search the wrong space entirely.
What stops the bot answering from irrelevant chunks?
A relevance floor. Chunks whose similarity falls below it are never injected. Without that, an off-topic question still receives the least-irrelevant paragraphs on file and the model stretches them into a confident wrong answer.
Can I see what retrieval actually returned?
Yes. The playground runs a real query against a connection and shows the chunks, their scores, their source documents, and the latency. Retrieval on a production turn is visible in that turn's trace.
Does the answer cite its sources?
A grounded turn returns structured citations. The model is instructed to cite inline with [n] markers, and the sources are listed back to the reader, de-duplicated by document, so the same file is not named three times.
How do I keep the knowledge base fresh?
A connected source can sync on a schedule — realtime, hourly, daily, weekly, or only when you ask — and each source reports whether it is connected, syncing, synced, paused, or in error.
Ground the answer in something you can point at.
Connect the source, choose how it is cut, pick the retrieval your bot actually needs, and read the chunks it was handed before your customer does.