Skip to main content
The knowledge base turns approved organization content into searchable vector chunks. Administrators manage source metadata through the gateway; the agent performs asynchronous extraction, chunking, embedding, and Qdrant upserts.

​
Supported source types

Files

Upload PDF and DOCX documents through presigned object-storage URLs.

Text

Store authored text directly without an object-storage round trip.

URLs

Index one page or crawl related pages with depth and sync settings.

FAQs

Index question/answer pairs for retrieval and the high-confidence fast path.

​
Source lifecycle

Knowledge records can also be paused. They store title, description, catalog, source type, file/source metadata, sync policy, word and chunk counts, last indexed time, and an error message when processing fails.

​
Processing pipeline

1

Create source metadata

The authenticated administrator creates text/URL/FAQ metadata or requests a presigned file upload.
2

Queue ingestion

Confirmation or creation produces a document-ingestion BullMQ job with stable document and organization context.
3

Extract and normalize

Source-specific pipelines stream text from MinIO, authored content, a URL fetch/crawl, or FAQ pairs.
4

Chunk and embed

Content is divided into bounded units and embedded with retry and concurrency controls.
5

Upsert and report status

Vectors and payloads are written to Qdrant, then the agent reports counts and status to the gateway using AI_TOOL_SECRET.
Only index content you are authorized to process. Crawled pages and uploaded documents may contain personal data, secrets, or copyrighted material that should not be exposed to the model.
Last modified on July 17, 2026