liveBook · Manning

Grounding

As explained in Designing AI Systems

Grounding is the process of connecting an AI application's responses to authoritative external or organizational knowledge.

Grounding connects an AI application's responses to authoritative organizational knowledge instead of relying on plausible guesses. A Data Service turns policies, product documentation, support articles, technical manuals, and internal wikis into searchable indexes, then retrieves relevant passages for an application's question. section 5.0

A production grounding architecture separates responsibilities across a repeatable pipeline:

  • Ingest diverse documents through format-specific parsers, extract clean text and structure, and chunk content for retrieval. section 5.0
  • Store each team's knowledge in an isolated index with its own embedding model, chunking strategy, and metadata schema. section 5.7
  • Generate embeddings through the Model Service and abstract vector storage behind write and ranked-search operations. section 5.7
  • Combine semantic vector search with optional keyword search through Reciprocal Rank Fusion when lexical matching improves results. section 5.7
  • Replace old chunks when source documents change so indexes remain current without manual reprocessing. section 5.7

Searching an organizational index

A workflow can ask the platform for relevant information by passing a natural-language question and an isolated index such as "patient_procedures". The returned results provide ranked matches, scores, document identifiers, and metadata for grounding a response. section 5.0 section 5.7

relevant_info = platform.data.search(
    query=question,
    index="patient_procedures"
)
 

Use the search call to inspect the compact application-facing interface for retrieving relevant indexed knowledge. section 5.0

How Designing AI Systems covers it

How deep it goes
Explains ingestion, chunking, embeddings, indexed retrieval, storage abstraction, hybrid search, and document replacement, including pgvector internals.
Examples
Uses policies, product documentation, support articles, technical manuals, internal wikis, PDFs, HTML articles, and patient-procedure searches.
What it assumes
Assumes conversational memory as a separate capability and teaches unstructured-document retrieval through ingestion, chunking, and semantic matching.
What you can do after
Apply it via platform.data.ingest() and platform.data.search() to maintain indexes and retrieve ranked organizational knowledge.
How it connects
Ties grounding to the Data Service, Session Service, Model Service, embeddings, vector search, keyword search, and ingestion pipelines.

Also explained in

Compare how 13 books explain grounding →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Systems. The section links point to the parts of the book it draws on.