Document Stores

Lyntaris Document Stores are the central place to ingest sources, split into chunks, curate, and upsert into a vector store. They power retrieval for Agentflow V2.

Where to find Document Stores

Document Stores live under Vector Databases. Most stores are auto-created, and you only configure what you need.

  • Document Store tab: a default store named document_store is auto-created and used for most teams.
  • Q&A / Images / Videos tabs: curated stores are auto-generated and managed by the platform.
  • Custom Stores tab: create extra stores to isolate datasets by client, project, or domain.

Use the default store when you want a shared knowledge base. Use custom stores when you need separation or different embeddings/vector configs. The auto-generated stores cover most client needs; custom stores are for special cases or strict data boundaries.

Curated stores and manifests

Curated stores live in the shared prompting pipeline and are managed by the UI. They also drive autosync for FastAPI ingestion.

  • Q&A: shared/qanda/manifest.json and shared/qanda/InputJsonQandA.json are the content payloads. shared/qanda/metadata.json is the sync signal (touched on save).
  • Images: shared/vector-databases/images/metadata.json
  • Videos: shared/vector-databases/videos/metadata.json

When the metadata file changes, Lyntaris emits a notify event and FastAPI refreshes the corresponding ingest.

Core workflow

  1. Open a store and click Add Document Loader.
  2. Choose a loader (PDF, Web, API, etc) and fill in its required fields.
  3. Optionally select a Text Splitter to control chunk size and overlap.
  4. Click Preview to inspect chunking, then Process to create chunks.
  5. Use View & Edit Chunks to remove or fix any noisy content.
  6. Configure Upsert: pick Embeddings, Vector Store, and optional Record Manager.
  7. Click Upsert All Chunks.
  8. Use Retrieval Query to test results and tune retrieval settings.

For Q&A, Images, and Videos stores, most of the pipeline is already wired. You typically only add or edit entries and let Lyntaris handle the rest.

Loader actions

Each loader row includes:

  • Preview & Process: update loader settings and generate chunks.
  • View & Edit Chunks: inspect or delete specific chunks.
  • Upsert Chunks: upsert only that loader's chunks.
  • View API: shows the exact Upsert API call for this loader.
  • Delete: remove the loader and its chunks.

Store actions

From More Actions:

  • View & Edit Chunks: view all chunks across the store.
  • Upsert All Chunks: push the full store to the vector database.
  • Retrieval Query: test queries against the vector store (requires an upserted store).
  • Refresh: re-process all loaders and upsert again.
  • Delete: remove the store and all loaders.

Status lifecycle

Document Stores show a status that reflects ingestion and upsert state:

  • EMPTY: no chunks yet.
  • SYNC: chunks match the latest processed content.
  • STALE: loaders changed or processing is pending.
  • UPSERTING: upsert in progress.
  • UPSERTED: ready for retrieval in flows.

If a store is STALE, use Refresh to re-process and re-upsert.

Retrieval Playground

The Retrieval Query view lets you:

  • run test queries,
  • inspect retrieved chunks and metadata,
  • adjust vector store settings (for example, topK),
  • save the config back to the store.

This is the fastest way to validate that retrieval quality matches your expectations before connecting the store to a flow.

Using Document Stores in flows

After a store is UPSERTED, it becomes selectable in the builders:

  • Agentflow V2: use Knowledge (Document Stores).

Only upserted stores appear in the dropdowns.

API quickstart

Use the View API button on a loader to get a ready-to-run Upsert call. The endpoints are:

  • POST /api/v1/document-store/upsert/{storeId}
  • POST /api/v1/document-store/refresh/{storeId}

If your loader accepts file uploads, use FormData. For URL-based loaders, you can send JSON.

results matching ""

    No results matching ""