Knowledge graph

How production AI fits together

The methods we use, the services they belong to and the case studies that used them, drawn as a knowledge graph. Select a node to see what it connects to, or read the glossary below.

Select a node to see what it connects to. On a keyboard, Tab to a node and press Enter.

Glossary

Each definition links to the paper or primary source it rests on, where there is one.

Retrieval

Retrieval-augmented generation (RAG): A language model answers from passages retrieved at query time from an external index, instead of relying only on what it learned in training. Source: Lewis et al., 2020
Hybrid retrieval: Lexical search such as BM25 and dense vector search, fused by a weighted sum of their scores or by reciprocal rank fusion, because the two capture relevance in complementary ways. Source: Bruch et al., 2022
Reranking: A second, stronger model, such as a fine-tuned BERT, scores each candidate that a first-stage retriever such as BM25 returned, and reorders them before they reach the prompt. Source: Nogueira and Cho, 2019
Contextual chunking: Splitting documents into passages and adding the context each passage needs on its own, such as the company and period, so retrieval does not lose meaning. It cut top-20 retrieval failures by 49%. Source: Anthropic, 2024
Multi-hop retrieval: Finding and combining evidence from several documents when no single passage answers the question. HotpotQA was built to test it with about 113,000 such questions. Source: Yang et al., 2018

Knowledge graphs

Knowledge graph: A graph of data meant to gather and communicate knowledge about the real world: nodes are entities of interest and edges are the relations between them. A schema or ontology fixes what the types mean. Source: Hogan et al., 2021
GraphRAG: An LLM extracts entities and relations from the corpus, the graph is partitioned into communities and each community is summarised. The summaries answer questions about the whole corpus, where vector RAG struggles. Source: Edge et al., 2024
Personalised PageRank retrieval: A random walk that keeps restarting at the entities named in the question, so passages linked to several of them rank high even when their text looks unlike the question. HippoRAG uses it for multi-hop questions. Source: Gutiérrez et al., 2024
Ontology: A formal, shared vocabulary for a domain that defines each term through its relationships to other terms; in OWL 2 it is written as classes, properties, individuals and data values. A knowledge graph’s schema follows it, so the same fact is stored the same way. Source: W3C, OWL 2 Overview
Entity resolution: Identifying the descriptions, in one source or across several, that refer to the same real-world entity when no shared identifier links them, and merging them into one node. A wrong merge creates a path that does not exist; a missed one breaks a path that does. Source: Christophides et al., 2020
Bitemporal data: Facts stored with two times: when they hold in the world (valid time) and when the database recorded them (transaction time). SQL:2011 calls tables with both periods bitemporal; they answer what was true, and what was known, on any date. Source: Kulkarni and Michels, 2012

Agents

AI agent: A language model that interleaves reasoning with actions, such as calls to tools, APIs or knowledge bases, and uses what those actions return to decide its next step. Source: Yao et al., 2022
Model Context Protocol (MCP): An open standard for connecting AI applications to external systems such as data sources, tools and workflows, introduced by Anthropic in November 2024. An integration built once can be shared by many agents. Source: modelcontextprotocol.io
Agent memory: The state an agent keeps across steps and sessions: business rules, history and earlier decisions. Zep stores it as a temporal knowledge graph that invalidates contradicted facts instead of deleting them, so it can answer what was true when. Source: Rasmussen et al., 2025
Human-in-the-loop approval: A checkpoint where a person approves, edits or rejects an action an agent proposes before it takes effect, placed where an error would be costly.

Evaluation

LLM evaluation: Measuring outputs against agreed criteria on a fixed set of real cases, so each change can be compared with the last. For RAG, retrieval and generation are scored separately. Source: Yu et al., 2024
Faithfulness: The share of an answer’s statements that the retrieved context supports. RAGAS splits the answer into statements and checks each one against the context. Source: Es et al., 2023
LLM-as-a-judge: A language model scores or compares outputs against a rubric. GPT-4 agreed with experts on 85% of votes, against 81% between experts, but judges show position and verbosity bias, so they are calibrated against people. Source: Zheng et al., 2023
Simulated users: AI-driven test users that rephrase, interrupt and change intent the way real callers and chat users do, so a conversational system is tested before each release.

Data

Data pipelines and quality checks: Batch and streaming flows that bring source data into the warehouse, index or graph, with tests and data contracts, so AI systems only see data that passed them.
Analytics platforms: Warehouses, lakehouses and real-time stores modelled for the questions people and AI systems ask, at a predictable cost.

Weighing vectors, a graph or both for your own data? We test it on your questions first.