Skip to main content
This page is for people who want to know what happens to a document after it goes into a knowledge base: where Enconvo keeps it, how it splits and indexes it, and how search picks results. Use it to tune a collection, fix poor results, or call knowledge bases from your own code. If you are new to knowledge bases, start with Knowledge Base. In short: Enconvo reads each item’s text, splits it into chunks, turns each chunk into a vector, and stores the chunks in a local LanceDB database. A search looks for chunks by meaning and by exact words at the same time, then blends the two.

​
Where the data lives

All knowledge bases share the one LanceDB database, and each knowledge base is one table in it. When Enconvo rebuilds a knowledge base’s index, it writes a new table beside the old one and switches over when the new one is ready, so search keeps working during the rebuild. Each knowledge base has an id. The APIs take it as knowledge_base_id, and Enconvo runs the knowledge base as the command knowledge_base|<id>. To copy a knowledge base’s id, open Knowledge Base Settings from its header and go to Advanced. Storage depends on the original assets, extracted text, chunk count, and embedding dimensions. Monitor actual disk use for your collection instead of estimating it only from the number of documents. Keep source files and backups independently of the search index.

​
How content gets in

​
Reading

Enconvo reads each kind of item with the tool that fits it:

​
Splitting into chunks

Enconvo splits the text into chunks sized for the language it is written in: It splits on the most natural boundary that keeps a chunk within its size, trying in this order:
  1. Paragraphs
  2. Lines
  3. Sentences
  4. Clauses
  5. Words
  6. A hard cut (last resort)
Spreadsheets are split about one row per chunk, so a search for one row finds that row. These sizes are built in; there is no setting for them.

​
Embedding

Every knowledge base uses the same embedding model, EmbeddingGemma 300M. Choose where it runs in Settings → Memory → Embedding model:
  • This Mac: free, and nothing leaves your Mac, but it uses the GPU while new content is indexed. It needs a Mac with Apple silicon.
  • Enconvo Cloud: free with your Enconvo account. The text being indexed or searched is sent there.
Both run the same model, so switching keeps your search index as it is. Enconvo embeds chunks in batches.

​
Storing

Each chunk is one row in the knowledge base’s table: Enconvo also keeps a full-text index of text, built from short runs of characters, so it finds exact names, numbers, code and CJK words.

​
How search works

Each search runs up to three lookups at once over the knowledge base’s chunks:
  1. By meaning: the query is embedded with the same model, and the closest vectors are found.
  2. By words: a full-text (BM25) search for each word of the query, leaving out common English words.
  3. By title: items whose title matches the query, each with its chunk closest to the query.
Enconvo then blends them. Each list’s scores are scaled from 0 to 1, and a chunk’s score is 0.6 of its meaning score plus 0.4 of its word score. An item whose title matches, or a chunk that contains the whole query, ranks higher. Only a few chunks from any one item are kept, so one long document can’t crowd out the rest, and each chunk comes back with the text around it. There is no separate reranking step in knowledge base search.

​
Keep the index healthy

Enconvo creates the full-text index and tidies the tables on its own. It also rebuilds a knowledge base’s index when it was made with an older chunker or a different embedding model, from the text it already has, without reading the files again. To rebuild by hand:
  • The whole knowledge base: open Knowledge Base Settings from its header, go to Index & Search, and click Reindex… next to Rebuild the whole index. Searches keep working while it runs.
  • One item: right-click it and choose Reindex.
Knowledge Base Settings has five sections: General (name and description), Index & Search (health, items that need attention, the search model and reindexing), Sources & Capture (linked folders and capture tools), AI Clients (MCP), and Advanced (the id and deleting the knowledge base). Delete knowledge bases you no longer need to reclaim disk space.

​
Linked folders

For a missing item, inspect the source folder’s import and synchronization status first. In Sources & Capture, each linked folder shows whether it is syncing, when it last synced, or that its source folder can’t be found, with Sync Now, Relink… and Auto-sync. Then inspect the item’s content-extraction status and embedding status separately: the item’s Pipeline shows Content parsing and Search index as two stages, and readable extracted text does not guarantee that vector indexing has finished. Reindex the item after fixing its provider or file-access error. The source folder remains the authority for linked files. Keep offline-drive recovery separate from file deletion: when a source folder can’t be found, the knowledge base keeps its items. Review skipped-file counts before treating an import as complete.

​
Organize several knowledge bases

Each knowledge base has its own documents, index and table. Split your content in the way you search it: To search several at once, attach them with the context picker, or set the Agent’s attached knowledge bases. A search that names no knowledge base looks through all of yours except Memory.

​
Use knowledge bases from code

The Knowledge Base plugin serves its routes on the Local API, so other plugins, workflows and scripts can call them. The main ones: Open Settings → APIs for each route’s parameters on your installed version. Knowledge base steps fit into workflows and automations, for example:
For another AI client, use Knowledge Base MCP. It exposes selected knowledge bases through a local, read-only service; it does not export Agent Memory or recordings.

​
Troubleshooting

Knowledge Base Basics

Getting started with knowledge bases

Knowledge Base MCP

Search your knowledge bases from another AI client

Memory

The search model and what agents remember

Workflows

Automate knowledge base tasks