Vector Embeddings

Why Embeddings Matter for Role Memory

Role knowledge does not fit neatly into keyword searches. A successor asking "how did we handle the vendor renegotiation?" might find nothing with keyword search if the captured memory uses the word "contract" instead of "renegotiation". Semantic search — enabled by vector embeddings — resolves this. It finds entries that are semantically similar to the query, regardless of the specific words used.

Vector embeddings are the technical foundation of RolegacyAI's retrieval capability. Every memory entry is encoded as a high-dimensional vector that captures its semantic meaning. Retrieval queries are encoded in the same space, enabling the RAG Engine to find the most semantically relevant memories rather than the most lexically similar ones.

Embedding Models

RolegacyAI uses embedding models calibrated for enterprise knowledge content — dense retrieval models that perform well on the types of text generated in professional environments: decision records, process documentation, meeting notes, and operational communications. The specific model is configurable and can be updated as better models become available, with a re-embedding pipeline that can migrate the existing memory store to a new embedding space without loss of content.

Chunking Strategy

Memory entries are not embedded as entire documents. They are chunked into semantically coherent segments before embedding, with chunk boundaries designed to respect the natural structure of the content: paragraph breaks, list items, and section headings are preferred chunk boundaries over arbitrary character or token counts. Each chunk is embedded independently, and the embeddings for all chunks of a memory entry are indexed together, enabling retrieval at the chunk level with aggregation back to the entry level.

Role-Scoped Embedding Spaces

Embeddings are stored in role-scoped vector indexes rather than a single global index. This scoping ensures that retrieval is fast (smaller index), accurate (reduced noise from irrelevant entries in other roles), and access-controlled (retrieval from one role's index cannot surface content from another). Cross-role search — for authorised use cases like portfolio risk analysis — operates on a separate aggregated index with appropriate access controls.

Clustering and Topic Discovery

Beyond retrieval, embeddings enable clustering analysis on the role's memory store. Clustering surfaces thematic patterns — areas of the role where many related entries cluster together (high-activity knowledge domains) and areas where the embedding space is sparse (knowledge gaps). This topical clustering feeds into the coverage scoring and gap analysis systems, providing a data-driven signal about where the role's memory is concentrated and where it is thin.