“Proxy-Pointer RAG: Revolutionizing Knowledge Graph Efficiency by Eliminating Redundant Entity and Relation Extraction”

• Structure tree creation: We create a structure tree of a document that lists the hierarchy of sections as nodes, along with section title.
• Enrich and Adjust: We navigate through the tree rather than the text itself. We use the initial set of documents to refine and strengthen the index. Each section’s content is identified using line numbers. The section titles help locate the predicted yield index. Next, the LLM reviews all sections of the document, evaluating the actual yield index for every section based on extracted connections and entities. Any mismatches between the expected and real ratings are flagged for manual evaluation (for example, when the actual rating is “Low” but the index predicted “Medium”). The index classifications are updated according to feedback from human subject-matter experts.
• Route and Bypass: Once the above procedure is complete, we derive a thoroughly enhanced graphability index after reviewing several documents. From that point, high-yield sections (Very High, High, Medium) are sent to the LLM for thorough NER extraction. Low and Very Low sections are safely skipped.
• New Sections: Every document will include some sections absent from the index, which are flagged as Coverage Gaps. These require mandatory NER scans to avoid overlooking relevant connections. Once reviewed by human evaluators, commonly occurring sections can be added to the index, while distinctive sections such as Benchmark Replacement Setting can be disregarded.
• Reach stabilization. After only a handful of iterations, we anticipate prediction mismatches to decrease to nearly zero, and the number of “New Sections” to level off at around 20-25% (reflecting highly specialized or routine clauses), allowing the system to process vast document collections with a reliable balance of thoroughness and efficiency.

The graphability index should be maintained for each document type and may even need to be customized for specific large suppliers and partners, from whom we could receive hundreds of similar documents in a year.

Let’s now examine how this operates in a practical experiment.

The Experimental Setup

To validate this approach, I set up an experiment using three lengthy, publicly available Corporate Credit Agreements that I referenced earlier in my article on efficient Contract Comparison through Proxy-Pointer. These agreements originate from distinct companies (and industries), so their structure and formatting differ across documents.

1. Emerson Electric Co. (~228,000 characters)
2. AT&T Inc. (~214,000 characters)
3. Texas Roadhouse, Inc. (TRoadhouse) (~434,000 characters)

Baseline Graphability Index

Our objective is to develop and iteratively confirm a predictive Graphability Index. We begin with a foundational baseline index mapping typical credit agreement sections to their anticipated relational density:

{
  "document_type": "credit_agreement",
  "very_high_graphability": [
    "Litigation",
    "Environmental Matters",
    "Subsidiaries",
    "Payment of Obligations",
    "Maintenance of Property",
    "Mergers and Sales of Assets",
    "Commitment Schedule",
    "Sanctions and Anti-Corruption",
    "Designation of Subsidiary Borrowers",
    "Definitions",
    "Events of Default",
    "Successors and Assigns"
  ],
  "high_graphability": [
    "Company Guarantee",
    "The Facility",
    "Facility Letters of Credit",
    "Corporate Existence and Power",
    "Corporate Authorization",
    "Financial Information",
    "Compliance with Laws",
    "Use of Proceeds",
    "Arranger and Syndication Agent",
    "Eurocurrency Payment Offices",
    "Defaulting Lenders"
  ],
  "medium_graphability": [
    "Swing Line Loans",
    "Competitive Bid Advances",
    "Credit Extensions",
    "Designation of a Subsidiary Borrower",
    "Successor Agent",
    "Funding Indemnification",
    "Acceleration and Collateral Accounts",
    "Collateral"
  ],
  "low_graphability": [
    "Accounting Terms",
    "Interest Rate Changes",
    "Method of Payment",
    "Telephonic Notices",
    "Market Disruption",
    "Judgment Currency",
    "Change in Circumstances",
    "Confidentiality"
  ],
  "very_low_graphability": [
    "No Waivers",
    "Counterparts and Integration",
    "Governing Law",
    "Waiver of Jury Trial",
    "No Fiduciary Duty",
    "Service of Process",
    "Miscellaneous",
    "Electronic Communications",
    "Exhibit",
    "Table of Contents"
  ]
}

The process is divided into three phases. First, the Emerson agreement is processed to determine the initial savings. Any uncovered general sections (deltas) identified in Emerson are incorporated into the index. Then the updated index is applied to AT&T, with any final edge cases added to the index if needed. Finally, the fully refined index is tested against the large TRoadhouse agreement to measure the total improvement. The objective is that by the time the TRoadhouse agreement is reviewed, mismatches should be considerably lower than in the prior two, as the index stabilizes.

Evaluation Criteria

For every section, we compare the index-predicted graphability with the actual rating determined by the LLM based on the relations and entities found. In the report, results are organized into three groups:

Perfect Alignment: The index accurately forecast the section’s graphability rating.

Minor Deviations: The index predicted a yield (e.g., Medium) that slightly varied from the manual assessment (e.g., Low).

Coverage Gaps / New Sections: The section was unique to the document and not yet included in our predictive index.

Results & Iterative Enrichment

Let’s proceed with Phase 1 — Emerson

Phase 1: Emerson Credit Agreement (Testing the Baseline)

We processed the 95 sections of this agreement using the baseline index. In this initial run, 66 out of 95 sections (70.0%) matched perfectly. The index accurately mapped standard provisions like “Mergers and Sales of Assets” as highly graphable, while correctly labeling “Accounting Terms” and standard boilerplate Exhibits as low-yield. No mismatches occurred between actual and predicted ratings from the index.

However, 29 sections (~30%) were flagged as New Sections, identified as Coverage Gaps. Upon review, while many were highly specialized administrative clauses (such as “Ratable advances” and “Notification of advances”) and appropriately left as gaps, several standard sections (like “Types of Advances,” “Compliance with ERISA,” and “Interest Payment Dates; Interest and Fee Basis”) needed to be added to the index. Based on their assessed actual yield, I added these specific clauses to the “Medium” and “Low” tiers of the graphability index, enriching the baseline for the subsequent phase.

A key finding is that even with this initial baseline index, 36,880 characters of text classified as “Low” or “Very Low” yield were correctly identified as noise by the index. As a result, skipping these sections and not sending them to the LLM for processing can lead to a 16.10% decrease in the overall LLM processing workload.

The following data highlights the match quality and the efficiency of yield prediction:

Here is a small sample of rows from the base table used for section-by-section comparison:

Node ID	Section Header	Approx. Chars	Entities (Est.)	Relations (Est.)	Actual Rating	Predicted Rating (Index Match)	Match Quality
0002	Section 1.01 Definitions	44,400	252	402	Very High	Very High (Definitions)	🟢
0003	Section 1.02 Accounting Terms and Determinations	320	4	4	Low	Low (Accounting Terms)	🟢
0004	Section 1.03 Types of Advances	800	19	2	Low	New Section	⚪
0006	Section 2.01 The Facility	2,320	27	21	High	High (The Facility)	🟢
0007	Section 2.02 Ratable Advances	3,840	56	19	Very High	New Section	⚪

The following are a few examples of extractions:

- **Company Guarantee (Very High)**:
  - *Entities*: Guarantor, Agent, Obligations
  - *Relations*: [Guarantor]-(guarantees)->[Obligations], [Guarantor]-(indemnifies)->[Agent]
- **Mergers and Sales of Assets (Very High)**:
  - *Entities*: Borrower, Assets, Buyer
  - *Relations*: [Borrower]-(sells)->[Assets], [Borrower]-(merges_with)->[Buyer]
- **Ratable Advances (Very High)**:
  - *Entities*: Advance, Lender, Borrower
  - *Relations*: [Lender]-(makes)->[Advance], [Borrower]-(receives)->[Advance]
- **Method of Payment (Low)**:
  - *Entities*: Agent, Accounts, Funds
  - *Relations*: None (section contains purely administrative procedures with minimal relevant relational links)

Phase 2: AT&T Credit Agreement (Refinement Stage)

Next, we applied the enhanced index to the AT&T Credit Agreement. The document consists of 77 sections with approximately 214,000 characters in total.

The results showed notable progress. 55 out of 77 sections (71.4%) achieved Perfect Alignment, nearly identical to the Emerson results. There were also 4 mismatched sections where the actual and predicted graphability ratings differed, accounting for only 5%. No adjustments were made to the index to prevent overfitting to individual documents. A reduction in Coverage Gaps was observed as only 18 sections (23.4%) had them, improving from Emerson’s 30%. All such sections were judged to be procedural noise from a knowledge graph perspective—examples include calculations of time periods, extensions of termination dates, or subordination clauses. These are considered low-yield segments in NER terms and should be excluded from future LLM scanning. However, to verify the robustness of the current index, these were not included during testing against the TRoadhouse document.

The potential savings in LLM usage grew significantly. Since the index could reliably detect extensive parts of the document as low-yield content—such as interest rate calculations, increased costs clauses, along with Table of Contents and trailing Exhibits—the system marked 72,763 characters as unnecessary for scanning. Implementing this approach in production could deliver a 33.94% reduction in processing requirements, while still ensuring all high-value relational information is captured.

The match quality and efficiency results are summarized below:

A small excerpt from the section rating analysis table is shown below:

Node ID	Section Header	Approx. Chars	Entities (Est.)	Relations (Est.)	Actual Rating	Predicted Rating (Index Match)	Match Quality
0017	SECTION 2.12. Payments and Computations	1,520	21	5	Low	Low (Payments and Computations)	🟢
0018	SECTION 2.13. Taxes	3,360	14	10	Medium	Medium (Taxes)	🟢
0019	SECTION 2.14. Sharing of Payments, Etc.	800	8	6	Low	Low (Sharing of Payments)	🟢
0020	SECTION 2.15. Evidence of Debt	640	10	2	Low	Low (Evidence of Debt)	🟢
0021	SECTION 2.16. Use of Proceeds	320	8	4	High	High (Use of Proceeds)	🟢
0022	SECTION 2.17. Increase in the Aggregate Commitments	2,800	22	9	Medium	New Section	⚪
0023	SECTION 2.18. Extension of Termination Date	3,120	20	25	Medium	New Section	⚪
0024	SECTION 2.20. Replacement of Lenders	1,920	19	12	Medium	Medium (Replacement of Lenders)	🟢
0025	SECTION 2.21. Benchmark Replacement Setting	12,560	61	31	High	High (Benchmark Replacement Setting)	🟢

Here are a selection of extraction examples:

- **Certain Defined Terms (Very High)**:
  - *Entities*: Base Rate, Margin, SOFR
  - *Relations*: IS_A, PART_OF, CONTROLS, ROLE_OF, REFERENCES (defining these terms establishes the core ontology, supports standard entity normalization, and enhances semantic structure)
- **Conditions Precedent (Medium)**:
  - *Entities*: Closing Date, Certificates, Approvals
  - *Relations*: [Lender]-(requires)->[Certificates], [Agent]-(receives)->[Approvals]
- **Accounting Terms; Interpretive Provisions (Low)**:
  - *Entities*: GAAP, Accounting Principles
  - *Relations*: None (this section consists entirely of administrative and interpretive content, with little to no significant relational data)

Phase 3: TRoadhouse Credit Agreement (Final Validation)

Although only the first document was used to build and refine the graphability index, let’s now validate it against the TRoadhouse credit agreement. Before proceeding, it is important to note that

It’s important to consider several differences, not just between the documents themselves, but also across the domain and industry. Emerson and AT&T are large, blue-chip utility and telecommunications providers, while Texas Roadhouse is a mid-sized restaurant chain. The agreements for Emerson and AT&T read like sovereign corporate treasury documents shaped by credit agency ratings, whereas Texas Roadhouse’s agreement is heavily customized and built specifically around restaurant lease arrangements. In terms of size, this document contains around 434,000 characters, making it nearly as large as the previous two combined, with more than 100 sections in its structure tree.

Put simply, if the Graphability Index performs well on this document, the idea that document structure can reliably predict the yield of entities and relationships will be proven beyond any doubt.

And here are the results — the index delivered outstanding performance. 81 out of 102 sections (79.4%) matched the index’s predictions exactly. There were zero cases where the actual rating diverged from the predicted one. The model perfectly identified high-yield sections such as “Letters of Credit” and standard “Affirmative and Negative Covenants,” which should trigger full extraction. The remaining 21 sections (20.6%), categorized as coverage gaps, consisted of low-yield administrative clauses (e.g., Rounding, Erroneous Payments) and procedural noise (e.g., Divisions, Commitments).

However, the real value emerged in payload efficiency. Several additional low-yield sections — including Accounting Terms, Rounding, Administrative Agent, and Miscellaneous — were identified beyond the Exhibits. The Schedules were evaluated individually based on their content value. While certain schedules like Liens and Investments matched the index’s “High” rating, others such as Existing LCs were classified as gaps.

The combined Low and Very Low categories translate to a net savings of 38% by following the predictions and skipping those sections entirely. This confirms the practical viability of the approach.

Below is the yield processing efficiency table:

Here are some examples of section ratings:

Node ID	Section Header	Approx. Chars	Entities (Est.)	Relations (Est.)	Actual Rating	Predicted Rating (Index Match)	Match Quality
0104	7.14 Financial Covenants	720	12	1	Very High	Very High (Financial Covenant)	🟢
0105	8.01 Events of Default	3,200	30	21	Medium	Medium (Events of Default)	🟢
0108	Article 9: ADMINISTRATIVE AGENT (Aggregated)	4,880	2	0	Low	Low (Duties of Agent)	🟢
0119	Article 10: MISCELLANEOUS (Aggregated)	18,000	2	0	Very Low	Very Low (Miscellaneous)	🟢
0144	Schedule 2.01A Commitments	4,000	2	0	Very High	Very High (Commitment Schedule)	🟢
0145	Schedule 2.01B L/C Commitments	2,000	2	0	Very Low	New Section	⚪
0146	Schedule 2.03 Existing L/Cs	3,000	3	0	Very Low	New Section	⚪
0147	Schedule 5.01 Jurisdictions	6,000	2	0	Very Low	New Section	⚪
0159	Schedule 5.06 Litigation	5,000	2	5	Very High	Very High (Litigation)	🟢
0161	Schedule 5.09 Environmental	8,000	2	5	Very High	Very High (Environmental Matters)	🟢
0163	Schedule 5.13 Subsidiaries	40,000	2	5	Very High	Very High (Subsidiaries)	🟢

And here are a few examples of extractions:

- **Financial Covenants (Very High)**:
  - *Entities*: Borrower, Leverage Ratio, Fixed Charge Coverage Ratio
  - *Relations*: [Borrower]-(maintains)->[Leverage Ratio]
- **Investments & Liens (High)**:
  - *Entities*: Borrower, Lien, Property, Permitted Investments
  - *Relations*: [Borrower]-(grants)->[Lien], [Borrower]-(makes)->[Permitted Investments]
- **Defined Terms (Very High)**:
  - *Entities*: Adjusted Term SOFR, Base Rate, Defaulting Lender
  - *Relations*: IS_A, PART_OF, CONTROLS, ROLE_OF, REFERENCES (Definitions form the ontology backbone, creating canonical entity normalization and robust semantic inheritance)

Conclusion

Today’s Knowledge Graph pipelines are fundamentally inefficient. We force costly LLMs to process entire enterprise corpora, even though only a small fraction of those documents contain meaningful relational intelligence.

This article demonstrated that document structure alone can act as a reliable predictor of graph extraction yield.

By integrating Proxy-Pointer’s structural analysis with Graphability Indexing, we can transition Knowledge Graph ingestion from brute-force semantic scanning to targeted structural routing. Rather than repeatedly processing full 500,000-character agreements, the system identifies which regions of a document family consistently produce valuable entities and relationships — and which are mostly boilerplate noise. We can simply bypass the noise entirely, without resorting to workarounds like smaller LLMs to cut costs.

Across three large real-world credit agreements spanning different industries, the index stabilized quickly after just a few iterations and consistently achieved significant payload reductions while maintaining high-value relational extraction.

More importantly, this calls for a shift in how we think about extraction architecture. Rather than treating documents as flat text streams, Proxy-Pointer treats them as structured semantic trees capable of predicting where meaningful knowledge is likely to reside before extraction even begins.

As enterprise GraphRAG systems scale across millions of contracts, filings, policies, and agreements, this kind of structure-aware ingestion could play a key role in making large-scale Knowledge Graph construction operationally sustainable.

Open-Source Repository

Proxy-Pointer is fully open-source (MIT License) and available at the Proxy-Pointer Github repository. You can install it with a single pip command using the package installer.

Clone the repo. Test it on your own documents. I’d love to hear your feedback.

Connect with me and share your thoughts at www.linkedin.com/in/partha-sarkar-lets-talk-AI

_{The credit agreements referenced here are publicly available at SEC.gov. Code and benchmark results are open-source under the MIT License. Images in this article were generated using Google Gemini.}