Creating a "company brain" for LLMs, which reliably uses a company's scattered knowledge, is far more complex than a simple demo project. A true context layer is a sophisticated system that continuously maps diverse data sources into normalized, typed context items. It indexes these items using multiple specialized methods and composes the most relevant subset for a given query, all while managing tenancy, permissions, freshness, and cost.

The core challenge lies in continuous mapping. Ingestion isn't a one-time batch job but an ongoing reconciliation loop, essential because data sources are constantly changing. Building an ontology, rather than a collection of individual connectors, simplifies this process. Each source is declaratively described, allowing a single mining engine to handle various data warehouses and wikis. Idempotency is achieved by deriving item identity from tenant, type, and path, enabling retries and crash recovery. Deletion is particularly complex, with multiple channels through which items can disappear, requiring careful handling to avoid data inaccuracies.

Indexing requires multiple approaches beyond a single vector database. A relational store serves as the source of truth, with other indexes like keyword and vector indexes acting as rebuildable projections. A knowledge graph is crucial for understanding structural relationships. Retrieval is an orchestration problem, not just a simple top-k retrieval. It involves composing dynamic strategies based on the query's scope and context, ensuring that the LLM accesses the most pertinent information from various sources.