Embedding models

ContextSage Pro unlocks larger embedding models via the bridge. Choosing the right model directly affects how well documents are retrieved — not just how fast they load.

Available models

Model	Context window	Size	Best for
MiniLM L6-v2	256 tokens	80 MB	Quick setup, short documents
BGE Large EN v1.5	512 tokens	1.3 GB	Legal documents, contracts, long clauses ✅
Nomic Embed	8192 tokens	270 MB	Very long documents, research papers
mxbai-embed-large	512 tokens	1.3 GB	General purpose, strong retrieval

Why context window matters for legal documents

The embedding model converts each chunk of text into a vector. If a chunk is longer than the model’s context window, the tail is silently truncated — the vector only represents the first N tokens, and the rest of the chunk is invisible to retrieval.

For legal contracts, this is a critical issue. A non-compete clause or limitation of liability paragraph typically runs 300–500 tokens. With MiniLM’s 256-token window, more than half the clause is cut off before embedding.

Measured on a real contract (Quaker non-compete agreement, 8 pages)

Model	Context window	Chunks over window	Result
MiniLM L6-v2	256 tokens	11 / 16 (69%)	Most clauses partially embedded
BGE Large EN v1.5	512 tokens	0 / 16 (0%)	Every clause fully embedded ✅

With MiniLM, 11 of 16 chunks exceeded the 256-token limit. The non-compete restriction itself (chunk 5, 484 tokens) — the most important clause in the document — had nearly half its content truncated during embedding. Retrieval on the truncated vector would return the chunk, but the model would only see an incomplete clause.

BGE Large embeds all 16 chunks in full. The same document, zero truncation.

Tip

If you are working with legal contracts, employment agreements, or any document with long structured clauses, use BGE Large EN v1.5. MiniLM is fine for short memos or quick notes.

Chunk size calibration

The chunk sizes ContextSage uses are calibrated to match the active embedding model’s context window. Switching models re-calibrates automatically:

Model	Default chunk size	Max chunk size
MiniLM L6-v2	200 tokens	256 tokens
BGE Large EN v1.5	400 tokens	512 tokens

This means switching to BGE Large produces fewer, larger chunks per document — each chunk contains a complete legal clause rather than a mid-sentence fragment. A typical 8-page contract produces ~16 well-formed chunks with BGE Large vs ~20 smaller fragments with MiniLM.

How to switch models

1. Make sure the bridge is running (pnpm bridge)
2. Open Settings → Embedding model in the app
3. Select your model — the bridge downloads it on first use
4. Re-ingest any documents you uploaded under the previous model

Caution

Documents ingested under one model must be re-ingested after switching. Vectors from different models are not comparable — mixing them in the same index produces incorrect retrieval results.