Infrastructure for Semantic Search Across Knowledge Base

Semantic search requires that documents exist in a findable repository with at least basic metadata — deliverable type, practice area, project context. Methodology documentation, template standards, and knowledge sharing processes exist at L2, meaning documents are at least deposited into repositories with some structure. However, the gap between documented methodology and actual practice means search results surface official artifacts that may not reflect how work is actually done, reducing trust in retrieved results.

Semantic search can only retrieve documents that were captured into the indexed repository. Mandated upload practices mean deliverables are deposited, creating an indexable corpus. However, the contextual metadata that enables semantic relevance — why this approach was chosen, what problem it solved, what alternatives were rejected — is rarely captured. The AI can match query terms to document content but cannot reliably assess relevance to the consultant's actual need without rich context capture.

Semantic search requires consistent metadata schema to filter and rank results by industry, service line, deliverable type, and project phase. The ps-km baseline has taxonomy defined — industry codes, service line classification, deliverable type taxonomy, folder hierarchies. This consistent schema enables the AI to scope searches ("find strategy deliverables for healthcare clients in the assessment phase") rather than returning undifferentiated full-corpus results. Without schema, semantic similarity alone produces noisy, poorly-filtered result sets.

Semantic search requires API access to the document repositories for indexing and query-time retrieval. The baseline indicates web-based repositories with search functionality exist, and modern SharePoint/Confluence installations expose APIs sufficient for content ingestion. The search engine can crawl and index document content and metadata. However, content locked in binary formats (docx, pptx) requires extraction pipelines, and legacy system constraints limit real-time index updates.

Semantic search utility degrades as stale documents accumulate in the index without purging. The baseline confirms case studies are never refreshed and search returns results from 5+ years ago. At L2, the index grows unchecked — each query must sift through outdated methodology documents, deprecated templates, and superseded case studies alongside current content. Relevance ranking is polluted by content that hasn't been reviewed or retired. Consultants learn to filter by date manually, partially defeating the semantic relevance value.

Semantic search across the knowledge base primarily needs access to document repositories — a point integration sufficient for the core use case. The search engine indexes available repositories and serves results. Deeper integration with CRM (to surface knowledge relevant to specific client contexts) or PSA (to suggest documents based on current project phase) would enhance relevance but isn't required for the capability to function. The knowledge base remains largely standalone from CRM and delivery systems.