Grounded retrieval. Persistent memory. Source-of-truth attribution.
Hybrid retrieval combining dense vector search, lexical search, and reranking. A multi-tier memory model retains what matters across turns and runs with document-level access control and source attribution.
Why hybrid retrieval, not just vector search.
Pure vector search misses literal terms such as account numbers, regulation IDs, and product SKUs. Pure lexical search misses semantic intent. Nexoraa runs both, then reranks the union with an LLM-based or cross-encoder reranker.
The result is higher precision on enterprise content where exact identifiers and conceptual meaning both matter.
Three tiers of memory.
Working memory
In-execution context. Holds intermediate values, prior steps, and current sub-task state. Cleared at workflow completion.
Episodic memory
Retains historical runs and outcomes. Used by agents to reproduce decisions, avoid repeated errors, and learn from prior outcomes under retention policy.
Entity memory
Structured business data such as customers, accounts, contracts, claims, and vendors. Queryable through governed tools, not free-form access.
Retrieval is governed before content reaches an agent.
Document classification
Public, internal, confidential, and restricted classifications are applied at ingestion.
Document-level access control
Specific documents are restricted by role and enforced at retrieval, not after.
Secure RAG authentication
Retrieval requires user authentication; queries are scoped to the requesting user's role.
Source attribution
Every retrieved chunk returns its source document, page, and chunk identifier, surfaced in the agent's output.
Stale data detection
Sources beyond defined refresh thresholds trigger alerts and may block retrieval.
Right to be forgotten
Configurable retention with automated purge of embeddings and source artifacts.
Enterprise content enters through a controlled pipeline.
Multi-format support
PDF, DOCX, XLSX, PPTX, HTML, Markdown, and scanned images with OCR.
Preprocessing
Cleaning, deduplication, language detection, and metadata extraction before embedding.
Chunking strategy
Configurable size, overlap, and strategy by document type with A/B testing.
Incremental re-indexing
Detect changed, new, and deleted documents; update embeddings incrementally.
Embedding versioning
Track which embedding model was used and manage migration when models are upgraded.
Automated ingestion
Scheduled or event-driven ingestion from SharePoint, S3, file shares, and databases.
Retrieval is monitored continuously.
Relevance scores, citation rates, empty-retrieval rates, and chunk hit rates are surfaced in dashboards. Ground-truth test sets validate retrieval quality before knowledge changes are promoted to production.