A knowledge graph deserves a place in enterprise knowledge-base selection when valuable questions repeatedly require connecting several objects, and the identity, direction, validity and provenance of those relationships can be maintained. More nodes and edges do not establish more reliable answers. Technology and knowledge-management leaders should first decide which relationships merit governance, which questions should use them, and which conclusions still require original evidence.
Separate three question types before choosing an architecture
Retrieval-augmented generation, or RAG, finds evidence before a model composes an answer. A knowledge graph represents identified objects and relationships, such as a supplier providing a component used in a product. A graph can connect scattered facts. Using a dedicated graph database, extracting relationships with a model, and generating community summaries are nevertheless three separate decisions that a proposal should explain individually.
The first question type is local fact lookup, such as a maintenance procedure for one model. The second traces relationships, such as which current products are connected to a supplier's disruption. The third synthesizes a corpus, such as recurring themes across years of failure reports. The first prioritizes retrieval and versions, the second complete paths, and the third broad coverage without erasing differences. Richer prose is not a shared acceptance criterion.
Microsoft GraphRAG documentation distinguishes local search combining entities and source text from global search using community reports.[1] That mechanism supports routing by question type. It does not establish that every implementation called GraphRAG uses the same algorithm or is suitable for an exact procurement-impact list.
The earlier long-context and RAG selection framework addresses material scope and reading cost. The additional decision here is whether the required relationships already exist in operational systems and have authoritative owners. A controlled join between two stable tables may suffice. Automatically extracting a graph and then recovering facts that were already explicit can add avoidable work.
Extracted relationships and co-occurrence carry different evidential weight
The original GraphRAG paper organizes corpus-level questions around entities, relationships, communities and summaries. Its evaluation focused on sensemaking over two corpora and explicitly identified cross-domain generalization as an open issue.[2] This motivates a candidate architecture, not a transferable enterprise accuracy figure or an exemption from counterexample testing.
Each edge should identify both objects, relationship type and direction, effective start and end, original source and version, confirmation status, and an accountable owner. A document title alone is insufficient provenance. A reviewer should locate the supporting passage or business record and establish whether proposed cooperation or a cancelled agreement was mistaken for a currently valid relationship.
Microsoft's indexing-method documentation describes model-extracted relationships in Standard GraphRAG and text-unit co-occurrence relationships in FastGraphRAG.[3] Co-occurrence can support thematic exploration without proving supply, control or dependency. Define the intended use of each relationship before deciding which automated outputs may support factual queries.
Neo4j's VectorCypherRetriever illustrates similarity retrieval followed by traversal of existing relationships to expand context.[4] Retrieval and graph queries can therefore be combined without replacing all retrieval. Correct expansion still depends on the initial entity and the graph. Starting from the wrong similarly named supplier merely expands the wrong context faster.
Trace one supplier-impact list from records to an answer
The following constructed scenario compares implementation scope; it is neither a customer project nor a measured product result. At an evaluation date of November 10, a procurement manager asks which products in production require review if supplier S17 suspends supply. The output must identify candidate products, connecting components, evidence and missing information. It neither stops production automatically nor makes the procurement decision.
Use five groups of authorized records. R1 states that S17 currently supplies component C9. R2 connects C9 to current product P2. R3 connects C9 to P4 historically, but that relationship ended on October 31. R4 states that another supplier, S71, supplies C11, which is used in P3. R5 proposes that S17 supply C11, but has not been approved. S17 and S71 have similar names and distinct enterprise identifiers.
Resolve the supplier in the question to S17. Select confirmed supply relationships effective on the evaluation date, yielding C9, then traverse valid component-use relationships to P2 while retaining R1 and R2 as the evidence path. Exclude R3 because it expired, R4 because it concerns another supplier, and R5 because it is an unapproved proposal. The result should explain these different exclusions.
The supported conclusion is that P2 requires further review within the relationship scope included and checked. It is not that only P2 will stop production. Inventory, substitutes, lead times, alternative sources and completeness of the source inventory remain unproved. Identify any factory records not included. Failure to find a path is insufficient evidence that no exposure exists.
Add change record R6: S17's supply relationship for C9 ends on November 12 and S22 becomes the supplier. A November 10 query must retain the old path, while a November 13 query uses the new relationship. Overwriting the edge loses historical reproducibility; adding a new edge without closing the old one mixes dates. This case tests both modeling and continuing maintenance.
Compare three viable paths by the relationships worth maintaining
Path A retains existing hybrid retrieval with human verification of cross-document relationships. It fits a phase with few complex relationship questions, rapidly changing material, or no agreed entity identifiers. Correct model, version and retrieval failures first, and retain a human evidence table. Relationship checks continue to consume labor and are harder to reuse consistently, but deferring graph construction can be a defensible decision.
Path B maintains only relationships with decision value. Map authoritative supply and component-use records, effective dates and approval states, use documents for clauses and exceptions, and retrieve evidence along controlled paths. Joins can serve a fixed relationship structure. Compare graph storage when relationship variety or variable traversal depth makes it useful, rather than treating adoption of a graph database as the desired outcome.
Path C automatically extracts a graph and community information from a larger text collection for relationship discovery or corpus-wide synthesis. It fits work where relationships are principally embedded in text, broad connections are genuinely needed, and the team can handle extraction errors. Entity resolution, relationship review, rebuilding, authorization inheritance and version switching become additional dependencies. Open-source licensing does not remove their costs.
For the running example, if authoritative records already establish R1 through R4, testing B first is generally easier to validate than inferring those relationships from prose. This is a conditional engineering recommendation. Discovering potential supply-risk themes across years of discussion can justify a separate C exploration task, but its leads should not enter current supplier relationships without confirmation. Assign distinct outputs and owners to the two uses.
Calculate the investment threshold with maintenance in the same account
For a reproducible planning example, assume 400 monthly questions, of which 100 require relationship checking. The other 300 local questions retain the same baseline. Compare only the incremental relationship route; common costs cancel. Use a twelve-month planning period and an assumed loaded labor rate of CNY 180 per hour. Every effort, fee and improvement below is a budgeting assumption to replace with pilot records.
Under A, each relationship check takes twelve minutes: 100 × 12 ÷ 60 = 20 hours per month, costing CNY 3,600. B assumes 120 construction hours, allocated as ten hours per month, eight monthly maintenance hours, four review minutes per question or about 6.67 hours, and CNY 600 monthly operating expense. Using the exact fraction, its monthly total is (10 + 8 + 100 × 4 ÷ 60) × 180 + 600 = CNY 5,040.
C assumes eighty construction hours, eighteen monthly hours reviewing extractions and updates, six review minutes per question, and CNY 1,200 operating expense. Its monthly total is (80 ÷ 12 + 18 + 100 × 6 ÷ 60) × 180 + 1,200 = CNY 7,440. These fees are planning inputs, not vendor quotes. All routes must first meet identical correctness and coverage requirements; unverified quick answers cannot be counted as saved review effort.
At these assumptions, one hundred relationship questions do not justify B through labor savings alone. Let q be the monthly relationship-question count. A costs 36q yuan and B costs 3,840 + 12q. They are equal at q = 160. Only above that volume does B become cheaper under this simplified account. Equality provides no allowance for error risk, the time value of money or demand volatility.
At three hundred monthly relationship questions, A costs CNY 10,800, B CNY 7,440 and C CNY 11,040. If B's maintenance rises from eight to twenty-four hours, its fixed component becomes CNY 6,720 and its break-even volume rises to 280 questions. Frequent relationship changes consume reuse benefits. B also requires an upfront labor budget of CNY 21,600; spreading it across months does not remove the need to allocate those resources.
No production-loss reduction or procurement benefit is valued here because no actual consequence records are available. Infrequent but important questions may still warrant a limited graph. The business owner should provide auditable loss scenarios, alternative responses and evidence needs. Do not invent an incident value to make the project pay back or equate released hours with cash savings already realized.
Counterexamples can change the decision, not merely a score
The first is an incorrect identity merge. Combining S17 and S71 into one node wrongly adds P3 to S17's impact list. Inspect enterprise identifiers, alias mappings and relationship provenance. Split the entities, rebuild affected paths and retest both suppliers. If many objects lack distinguishable identity, improve master data before broad extraction amplifies the gap.
The second is a stale relationship. If R6 has taken effect but the old path remains in a current answer, source update, invalidation or serving-version activation has failed. Record source-change time, business effective time, index activation and query version. After repair, replay both dates to establish correct current behavior and preserved history. A last-updated timestamp alone does not demonstrate this.
The third is a summary that hides an exception. A community report may call supply stable while omitting an unapproved proposal or a decisive termination clause. Return to individual sources and separate thematic synthesis from object-level conclusions. Repair the summary or restrict it to navigation, then retest after removing a decisive clause and adding a contradictory one. Citing a report does not establish primary support for every assertion.
Test authorization changes through relationships and summaries as well. Permission to read one node does not imply permission to read all neighbors; generated summaries can indirectly expose restricted relationships. Submit the same question under different roles and inspect retrieval, model input and output. If the architecture cannot enforce the boundary, narrow or suspend that route rather than merely adding a warning to its answers.
A pilot should deliver relationship evidence and change records
For each question, fix entity identity, query date, role, required sources and allowed conclusion. The case's correct result set is P2, with three distinct exclusion reasons. Include missing-source, ambiguous-entity and no-valid-path questions and check that the system distinguishes their uncertainty. Answer counts, path counts and retrieved-chunk counts are not interchangeable.
Compare A, B and C on the same authorized snapshot and answer requirements. Record required-relationship coverage, unsupported edges, validity errors, complete evidence paths, correction minutes and update delay separately. A single preference score is inadequate. Reviewers should locate failure in identity mapping, extraction, path selection or answer generation instead of attributing every error to the model.
Choose pilot size from task variety and error consequences rather than declaring a fixed question count universally sufficient. Begin with a small relationship collection that humans can enumerate, then add authorized operational material. Keep tuning and acceptance questions separate and record maintenance effort by real update batch. A static demonstration without an R6-style change cannot establish continuing delivery capacity.
Confirm component support boundaries too. On November 16, 2026, the Microsoft GraphRAG repository describes a largely maintenance-mode project and cautions about indexing cost and the support status of its demonstration code.[5] That is a dated observation about this repository, not every graph product. Responsibility for versions, vulnerability fixes and replacement remains an enterprise decision.
Record a decision that can be repeated
The first decision-card field records question class, monthly volume, intended readers and error consequences. The second identifies authoritative entity records, relationship sources, validity, approval and maintenance ownership. The third documents existing-route failures and the particular missing relationship a graph would address. If the evidence only reveals a faulty retrieval field, fix it before broadening the project.
The fourth field lists construction hours, operating fees, monthly maintenance, per-question review time and assumption ranges for all three routes. The fifth retains inputs and expectations for normal cases, identity confusion, historical changes, missing evidence and permission changes. For this case, explicitly check P2, the S17/S71 distinction, dates around R6 and the prohibition on inferring production stoppage from a path alone.
Proceed when important questions gain evidence value, maintenance ownership is assigned, update and authorization tests pass, and conservative costs remain affordable. If only a few relationships qualify, deploy only those. If authoritative sources and continuing updates cannot yet be established, retain retrieval with human checks. The delivery unit is a maintainable, explainable class of answers, not a node count.
References
- [1] Microsoft GraphRAG — Query Engine Overview (accessed 2026-11-16)
- [2] Edge et al. — From Local to Global: A GraphRAG Approach to Query-Focused Summarization (arXiv v2, 2025-02-19)
- [3] Microsoft GraphRAG — Indexing Methods (accessed 2026-11-16)
- [4] Neo4j GraphRAG Python — User Guide: RAG, Vector Cypher Retriever (accessed 2026-11-16)
- [5] Microsoft GraphRAG — repository guidance and maintenance notice (accessed 2026-11-16)
