Embeddings¶

Placeholder — this page needs to be written.

Embeddings enable semantic search over the knowledge graph — finding entities by meaning rather than exact string match.

Where embeddings are generated¶

Embeddings are generated at the end of the ingestion pipeline, after entities are resolved and deduplicated. The EmbedderInterface takes a resolved entity record and returns a dense vector.

What gets embedded¶

• Entity name + description — the primary embedding, used for semantic search.
• Entity context — optionally, a snippet of the source text where the entity was first mentioned, to capture domain-specific usage.

Storage¶

Embeddings are stored in the bundle alongside entity records. At serve time, kgserver loads them into a vector index (currently in-memory; persistent index support is planned).

Matching embedding models¶

The model used at ingestion time must match the model used at query time. Mismatches produce meaningless similarity scores. The bundle manifest records the embedding model name and version.

Choosing a model¶

• For biomedical domains: PubMedBERT, BioLORD, or similar domain-adapted models outperform general-purpose models.
• For general domains: text-embedding-3-small or text-embedding-3-large (OpenAI) are practical defaults.
• For fully local deployments: Nomic Embed or similar open models.

See also¶

• Pipeline
• Querying with LLMs