The pipeline
From uploaded file to answered question
The full retrieval pipeline, managed.
One index per customer.
GTWY runs the pipeline. You run your product. Each of your customers gets their own index inside it — separate storage, separate retrieval scope, no shared pool to filter.
Ingest
What happens when a document is uploaded.
Parse
Format detected and text extracted with structure preserved — headings, tables and lists survive as boundaries rather than collapsing into one blob. Scanned PDFs and images go through OCR first.
Chunk
Recursive splitting on structural boundaries first, falling back to token windows. Default is roughly 800 tokens with ~15% overlap, so a sentence spanning a boundary is retrievable from either side.
Chunk size and overlap are configurable per platform account.
Embed
Chunks are embedded with the model you choose, on your provider key. Dimensionality follows the model — the index is built to match it, so switching models means a re-embed.
Index
Vectors and their source text land in that customer’s own index, tagged with the document and chunk they came from so retrieval can cite them.
Query
What happens when the agent needs to know something.
- 01RetrieveHybrid search — vector similarity and keyword matching over that customer’s index only
- 02RerankCandidates reordered by relevance to the actual question before any of them reach the model
- 03ReturnPassages handed to the agent with their source document and chunk, so answers can cite
All three happen inside the agent run. Retrieval is not billed separately and is not counted per query.
Freshness
When it becomes searchable, and when it stops being.
After an upload
Indexing starts as soon as the upload completes. A typical document is searchable within seconds; large files and OCR jobs take longer, and the status is visible to your user throughout.
After an update
Re-uploading a document replaces it. The old chunks are removed as the new ones land, so retrieval never returns both versions of the same passage.
After a deletion
Deletion removes the document and its chunks from the index. Once it is gone it cannot be retrieved, including by a run already in flight.
Synced sources
Connected sources re-check on a schedule you configure. Changed pages and files are re-ingested; removed ones are dropped from the index on the same pass.
Embedding models
Bring your own key.
Your embedding costs stay on your provider account, at your negotiated rate. We never mark them up because we never touch them.
Providers and models
The major hosted embedding providers, with the model chosen per platform account. The index is built to that model’s dimensionality.
Where the key lives
Encrypted at rest, used only to embed your customers’ content, never exposed in the embed or to your users. Rotate or revoke it at any time.
Changing model
A model change re-embeds every document, and those tokens bill to your provider account. You see an estimate before it starts — a re-index never runs without a cost preview.
If the key fails
An invalid, revoked or rate-limited key stops ingest rather than failing quietly. Documents stay queued, the failure surfaces in the panel and on your alerting, and ingest resumes once the key works — so your users never sit on a knowledge base that silently stopped updating.
Configuration
What you control as the platform.
Observability
What you see as the platform.
Query logs
What was asked, what was retrieved, and from which document.
Retrieval quality
Similarity scores and how often a query returns nothing usable.
Storage per customer
GB indexed per customer, against the cap you set.
Ingest failures
Which document, which stage, and why — parse, OCR, embed or index.
Index freshness
Last successful sync per connected source.
Embedding spend
Tokens embedded per period, so your provider bill is never a surprise.
API
If you’d rather not use the UI.
The embed gives your users a document manager. These do the same things from your backend.
curl --url 'https://db.gtwy.ai/api/rag/resource' -H 'authorization: <Bearer Token>' -H 'content-type: application/json' --data-raw '{
"title": "<Your Document Title>",
"description": "<Your Document Description>",
"settings": { "strategy": "recursive", "chunkSize": 4000 },
"collection_details": "<Your Collection Id>",
"url": "<Your Document URL>"
}'curl --url 'https://api.gtwy.ai/rag/query' -H 'authorization: <Bearer Token>' -H 'content-type: application/json' --data-raw '{
"query": "<Your Question>",
"resource_id": "<Your Resource Id>"
}'<script
id="rag-main-script"
embedToken="Add your embed token here"
src="https://chatbot.gtwy.ai/rag-prod.js"
parentId="Id of parent Container"
theme="dark/light"
defaultOpen="true/false"
></script>