Amin, Fakabbir · Oct 27, 2026
Beyond Similarity Search: Knowledge Graph Metadata Enrichment for Enterprise Retrieval-Augmented Generation
Markdown · JSON · Original source
Fakabbir Amin compares three retrieval designs for enterprise documents: text-to-SQL, dense-vector , and a two-stage GraphRAG design that uses a knowledge graph. The paper asks whether storing metadata and frequently changing information in a knowledge graph can improve retrieval over similarity search. On a snapshot of SEC EDGAR risk-factor filings, GraphRAG solved 16 of 20 questions on all three attempts, compared with 10 for text-to-SQL and 3 for vector RAG. The authors report a notable reversal for single-document questions: vector RAG passed 3 of 15, while GraphRAG passed 15 of 15. They attribute vector RAG’s result to company attribution not being recoverable from cosine similarity at this scale. The paper also reports that GraphRAG lost on structured aggregation queries. All six confident falsehoods in the came from text-to-SQL conflating people with the same surname; the authors say resolving identity to an identifier before reading text prevents that failure structurally. These results suggest, within this benchmark, that adding structured metadata may help retrieval when company or person identity matters, but may not suit every query type. The authors claim no advantage beyond this corpus, model, and question set. Would the reported differences persist on other corpora, models, or question sets?
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T19:17:30.608Z. First observed: 2026-10-30T06:42:39.255Z. Retrieved: 2026-10-31T10:18:53.753Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evidence is the abstract, not the full paper; it does not provide further methodological detail beyond the experiment description. The uses one SEC EDGAR snapshot, one model, and 20 questions across four categories, each asked three times; the authors explicitly claim no advantage beyond these conditions. Results were scored programmatically by an as judge, so the abstract does not establish how the ranking would compare with other evaluation methods.
Open question
Would the reported differences persist on other corpora, models, or question sets?
Source-stated
Amin compares text-to-SQL, dense-vector RAG, and a two-stage GraphRAG design using a knowledge graph for metadata and frequently changing information.
we evaluated three architectures: text-to-SQL over typed columns, dense-vector RAG with HNSW cosine search, and a two-stage GraphRAG design.
Source-stated
In the reported benchmark, GraphRAG solved 16 of 20 questions on all three attempts; text-to-SQL solved 10 and vector RAG solved 3.
GraphRAG solved 16 of 20 questions on all three attempts, compared to 10 for text-to-SQL and 3 for vector RAG.
Source-stated
For single-document questions, vector RAG passed 3 of 15 while GraphRAG passed 15 of 15; the authors attribute vector RAG’s result to company attribution not being recoverable from cosine similarity at this scale.
Vector RAG lost the single-document category it was expected to win, passing 3 of 15 where GraphRAG passed 15 of 15, because company attribution is not recoverable from cosine similarity at this scale.
Source-stated
GraphRAG lost on structured aggregation queries, and the authors report that all six confident falsehoods came from text-to-SQL conflating people with the same surname.
All six confident falsehoods in the benchmark were text-to-SQL conflating distinct people who share a surname
Source-stated
The authors limit their claim of advantage to the tested corpus, model, and question set.
We claim no advantage beyond this corpus, this model and this question set.
Escape Meridian interpretation
Within this benchmark, the results suggest structured metadata may help when company or person identity matters, but may not suit every query type.
All six confident falsehoods in the benchmark were text-to-SQL conflating distinct people who share a surname
Source and contributor context
Amin, Fakabbir · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Retrieval-augmented generation
A setup in which a model first looks up relevant documents and then answers using them.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: e905d714-f742-4ed0-89b3-f654ffaa196c.
Prepared by server-excerpt-repair/stream-excerpt-repair-v1 (original draft: gpt-6-luna); grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T19:17:30.608Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: a5af597228678c03737ad015ba1d6a08e3565a7401f23283bf96ea8d2e9c28b5. Source text hash: 3d6140a14782606f79d4d5fb282d46d68d46f21c6c3bdfecbd5596a24841bc07.
Revision and source-check history
Version 1 · 2026-11-10T19:17:30.608Z
Beyond Similarity Search: Knowledge Graph Metadata Enrichment for Enterprise Retrieval-Augmented Generation
Fakabbir Amin compares three retrieval designs for enterprise documents: text-to-SQL, dense-vector , and a two-stage GraphRAG design that uses a knowledge graph. The paper asks whether storing metadata and frequently changing information in a knowledge graph can improve retrieval over similarity search. On a snapshot of SEC EDGAR risk-factor filings, GraphRAG solved 16 of 20 questions on all three attempts, compared with 10 for text-to-SQL and 3 for vector RAG. The authors report a notable reversal for single-document questions: vector RAG passed 3 of 15, while GraphRAG passed 15 of 15. They attribute vector RAG’s result to company attribution not being recoverable from cosine similarity at this scale. The paper also reports that GraphRAG lost on structured aggregation queries. All six confident falsehoods in the came from text-to-SQL conflating people with the same surname; the authors say resolving identity to an identifier before reading text prevents that failure structurally. These results suggest, within this benchmark, that adding structured metadata may help retrieval when company or person identity matters, but may not suit every query type. The authors claim no advantage beyond this corpus, model, and question set. Would the reported differences persist on other corpora, models, or question sets?
Original source
cs.AI · Article
Source details
Original source date: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T19:17:30.608Z. First observed: 2026-10-30T06:42:39.255Z. Retrieved: 2026-10-31T10:18:53.753Z.
Limits
The evidence is the abstract, not the full paper; it does not provide further methodological detail beyond the experiment description. The uses one SEC EDGAR snapshot, one model, and 20 questions across four categories, each asked three times; the authors explicitly claim no advantage beyond these conditions. Results were scored programmatically by an as judge, so the abstract does not establish how the ranking would compare with other evaluation methods.
Open question
Would the reported differences persist on other corpora, models, or question sets?
Source-stated
Amin compares text-to-SQL, dense-vector RAG, and a two-stage GraphRAG design using a knowledge graph for metadata and frequently changing information.
we evaluated three architectures: text-to-SQL over typed columns, dense-vector RAG with HNSW cosine search, and a two-stage GraphRAG design.
Source-stated
In the reported benchmark, GraphRAG solved 16 of 20 questions on all three attempts; text-to-SQL solved 10 and vector RAG solved 3.
GraphRAG solved 16 of 20 questions on all three attempts, compared to 10 for text-to-SQL and 3 for vector RAG.
Source-stated
For single-document questions, vector RAG passed 3 of 15 while GraphRAG passed 15 of 15; the authors attribute vector RAG’s result to company attribution not being recoverable from cosine similarity at this scale.
Vector RAG lost the single-document category it was expected to win, passing 3 of 15 where GraphRAG passed 15 of 15, because company attribution is not recoverable from cosine similarity at this scale.
Source-stated
GraphRAG lost on structured aggregation queries, and the authors report that all six confident falsehoods came from text-to-SQL conflating people with the same surname.
All six confident falsehoods in the benchmark were text-to-SQL conflating distinct people who share a surname
Source-stated
The authors limit their claim of advantage to the tested corpus, model, and question set.
We claim no advantage beyond this corpus, this model and this question set.
Escape Meridian interpretation
Within this benchmark, the results suggest structured metadata may help when company or person identity matters, but may not suit every query type.
All six confident falsehoods in the benchmark were text-to-SQL conflating distinct people who share a surname
Source and contributor context
Amin, Fakabbir · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Retrieval-augmented generation
A setup in which a model first looks up relevant documents and then answers using them.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: e905d714-f742-4ed0-89b3-f654ffaa196c.
Prepared by server-excerpt-repair/stream-excerpt-repair-v1 (original draft: gpt-6-luna); grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T19:17:30.608Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: a5af597228678c03737ad015ba1d6a08e3565a7401f23283bf96ea8d2e9c28b5. Source text hash: 3d6140a14782606f79d4d5fb282d46d68d46f21c6c3bdfecbd5596a24841bc07.
uncertain · 2026-11-10T19:17:30.608Z · anthropic/claude-opus-5.5
Publication: e905d714-f742-4ed0-89b3-f654ffaa196c
attribution — pass: The paper is correctly attributed to Fakabbir Amin on arXiv. The draft later says "the authors" although the page lists a single author. This is a minor wording slip, not a misattribution; the smallest repair is "the author" or "Amin."
Authors:Fakabbir Amin
summary — uncertain: The figures match the abstract, but central terms are never explained for a non-specialist reader: RAG (not expanded in the body), text-to-SQL, dense-vector RAG, GraphRAG, knowledge graph, cosine similarity, SEC EDGAR and structured aggregation queries. "Notable reversal" also omits the source's point that vector RAG was expected to win that category. Smallest repair: add one plain-language line per core design, or use term-library tooltips, and say "the category it was expected to win."
Vector RAG lost the single-document category it was expected to win
date — pass: originalPublished 2026-10-27 matches the v1 submission date on arXiv. The later feed date is correctly not used.
[Submitted on 27 Oct 2026]
support — uncertain: The source-stated claims are accurate and the scope limit is preserved. The interpretive claim that "adding structured metadata may help" when identity matters is imprecise. The text-to-SQL system that made all six confident falsehoods also used structured typed columns. The source credits resolving identity to an identifier before reading text, not structured metadata as such. The inference is hedged but not labeled as Escape Meridian interpretation. Smallest repair: tie it to identifier-first identity resolution and label it as interpretation.
the failure that resolving identity to an identifier before reading any text prevents structurally; GraphRAG made none
media — pass: The selected media is the verified direct link to the original arXiv abstract page, matching the source URL. It is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A setup in which a model first looks up relevant documents and then answers using them.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA setup in which a model first looks up relevant documents and then answers using them.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition reference