Awesome Search KG

Enterprise Search

4 min read

Enterprise Search

Definition

Enterprise search is search applied to internal organizational content — documents, emails, files, messages, wikis, databases, and other knowledge assets — to help employees find information they need to do their work. Unlike e-commerce search, success is measured not by purchase conversion but by task completion, time-to-answer, and employee productivity.

Key Distinguishing Characteristics

No conversion rate. The canonical success signal of e-commerce (add-to-cart, purchase) doesn’t exist. Proxies include click-through, dwell time, zero-result rate, and user satisfaction surveys. See When There’s No Conversion Rate.

Authorization and permissions filtering. Results must be filtered based on the user’s access rights — a hard constraint absent from most consumer search. This makes pre-filtering and index segmentation critical design concerns.

Federated, multi-source retrieval. Content lives across many systems: SharePoint, Confluence, Slack, Google Drive, Jira, email. Enterprise search must query and merge results from heterogeneous sources, often with different schemas, freshness SLAs, and relevance models — the classic Federated Search problem (collection selection, query routing, results merging).

Structured knowledge graph integration. Organizations accumulate entity graphs (org charts, product catalogs, project hierarchies). Surfacing relationships — “who owns this project?”, “what documents are connected to this customer?” — requires knowledge graph integration.

RAG as default architecture. With the rise of LLMs, enterprise search increasingly powers retrieval-augmented generation (RAG) pipelines that synthesize answers rather than just returning a ranked list of links.

Core Challenges

  • Cold-start and low-query-volume: Internal query logs are sparse; hard to use implicit feedback.
  • Content freshness: Documents change frequently; indexing pipelines must be near-real-time.
  • Query ambiguity: Employees use internal jargon, project codenames, acronyms unknown to general language models.
  • Multimodal content: PDFs, slide decks, scanned documents, audio transcripts.
  • Privacy and compliance: GDPR, data residency, audit trails for what was retrieved and by whom.

Architectures

Classic lexical (BM25-based): Still dominant for structured intranet search; predictable, auditable.

Semantic / neural retrieval: Context-Aware Embeddings and Task-Aware Embeddings dramatically improve recall on paraphrastic queries that don’t match keywords. See How Context-Aware Embeddings Are Transforming Enterprise Search.

RAG pipeline: Retrieve relevant passages → rerank → synthesize answer with LLM. Dropbox Dash is a prominent example. See Using LLMs to Amplify Human Labeling and Improve Dash Search Relevance.

Hybrid search: BM25 + dense vector retrieval fused via RRF or linear combination, then reranked. Standard production pattern. See Hybrid Search Blueprint Series Semantic Boosting.

Unlike e-commerce, behavioral signals are weaker and noisier:

  • Recency: Newer documents usually more relevant than old ones.
  • Authority/provenance: Official policy docs > personal notes.
  • Org proximity: Content authored by or about the user’s team ranks higher.
  • Explicit feedback: Thumbs up/down ratings, bookmarks, re-queries.

LLM-as-judge can substitute for human labeling at scale when query volume is too low for implicit feedback loops. See LLM-as-a-Judge When to Use Reasoning CoT and Explanations.

Case Studies in the Vault

  • Dropbox — Dash product (enterprise search + RAG)
  • Slack — workplace message and file search
  • Elasticsearch / Elastic — widely deployed for internal search infrastructure

Graph View

Backlinks