From Prompt Rules to Golden Data Model: Production-Scale Generation of Personalized, Compliant Pharmaceutical Claims

Suyash Mishra; Qiang Li; Subhabrata Adak; Upender Phogat; Srikanth Patil; Satyanarayan Pati

cs.AIOct 22, 2026

From Prompt Rules to Golden Data Model: Production-Scale Generation of Personalized, Compliant Pharmaceutical Claims

Authors: Suyash Mishra, Qiang Li, Subhabrata Adak, Upender Phogat, Srikanth Patil, Satyanarayan Pati

Organizations: aRoche · bAccenture · cInvolead

Abstract

Large language models (LLMs) are increasingly used to generate domain-specific commercial & medical content (e.g., claims, statements, and promotional phrases) in regulated industries such as pharmaceuticals. However, ensuring its regulatory compliance, factual accuracy, and audience-specific personalization remains a major challenge. We report a production case study asking whether moving compliance and personalization signal from prompt-level rules into retrieval time structured context improves pharmaceutical claims generation, using Composer, a deployed internal tool built on a Golden Data Model (GDM) of 1,867 pharmaceutical claims spanning four products. The GDM integrates claim metadata, references, audience labels, material type, usage history, Match Text Variant (MTV) links, and a six dimension quality score (6C). Across six product-audience segments, the context pipeline improves intended indication (+29%), intended content type (+62%) and intended target audience (+67%), 6C quality score (+86%), and claim reuse (+34%). Production log analysis over the deployment period shows a 29.5% reduction in weekly API cost. We report this as descriptive deployment evidence. This work is a systems and data architecture contribution: it shows that in a regulated, high-stakes domain, restructuring claim data for retrieval-time filtering is a more scalable lever than expanding prompt-level rules, and documents the engineering path data cleaning, graph linking, and quality scoring.

Figures & tables

Figure 1: Business Solution Map for Composer Tool Compliant Claims Generation.

Figure 2: Comprehensive Functional Layer for Compliant Claims Generation.

Figure 3: Failure modes in legacy Veeva claim data. We observe two primary problem areas: ambiguous provenance, where the source of a claim is not always uniquely grounded, and heterogeneous data quality issues, including contradictory, duplicate, malformed, and low-value claims. These failure modes introduce retrieval ambiguity and reduce the effectiveness of downstream claim personalization.

Field	Type	Description
Claim ID	String	Unique claim identifier, e.g., CL-US-000754 .
Create Date	Datetime	Timestamp when the claim record was created.
Modified Date	Datetime	Timestamp when the claim record was last modified.
Product	String	Product associated with the claim, e.g., Ocrevus .
Country	String	Country or market where the claim is approved.
Match Text	Text	Original approved claim text. e.g., OCREVUS is proven effective in reducing relapses in RMS, slowing disability progression in RMS.

Table 1: Golden Data Model Schema for claim-level records. The schema integrates cleaned raw claim data with LLM-derived tagging layers and 6C quality scores. It also extracts tone signals and constraint-aware key phrases, enabling more controlled, compliant, and personalized downstream claim generation.

Product	#Claims	#Refs	#Materials	%Used	%With Ref	#MTV	#MTV Used	%MTV	%MTV Used
Hemlibra	745	721	593	80%	97%	297	187	39.87%	25.10%
Ocrevus	357	355	215	60%	99%	113	31	31.65 %	8.68 %
Ocrevus Zunovo	122	122	45	37%	100%	29	8	23.77%	6.56%
Vabysmo	643	640	489	76%	100%	358	102	55.68 %	15.86%
Total	1867	1838	1342	72%	98%	797	328	42.69 %	17.57 %

Table 2: Overview of claims, references, materials, and MTV (Match Text Variation, i.e. variant claims) availability and utilization across products. A total of 1,867 claims are identified across (Hemlibra, Ocrevus, Ocrevus Zunovo, Vabysmo) products for the US market. 72% of claims are utilized in materials, indicating strong adoption in content. But, for Ocrevus Zunovo it is only 37%. 98% of claims are Sourced from references, while the remaining 2% do not have any linked reference.

Metric	Pre	Post	$\boldsymbol{\Delta}$
Approved Claims	71%	100%	+29%
MTV Utilization	0%	34%	+34%
Indication Alignment	71%	100%	+29%
Content-Type Alignment	38%	100%	+62%
Target Audience Alignment	33%	100%	+67%

Table 3: Performance comparison between rule-based (Pre) and context-based (Post) systems. Here, we reports MTV utilization within the Pre/Post evaluation sample specifically.

Segment	Stage	C	U	Avg 6C score	Min	Max
OCR-MS-Nurse	Pre	5	2	0.83	0.83	0.83
	Post	5	5	0.91	0.83	1.00
OCR-MS-Patient	Pre	3	3	0.91	0.83	1.00
	Post	6	6	0.93	0.83	1.00
VAB-DME-Patient	Pre	2	0	0.00	0.00	0.00
	Post	3	3	0.86	0.83	0.89

Table 5: Quality vs. usage comparison before/after the Golden Data Model. C number of claims, U number of re-used claims

Input	Used		MTV		Ind		Type		Aud
	Pre	Post	Pre	Post	Pre	Post	Pre	Post	Pre	Post
VAB-RVO-HCP	80	100	0	20	20	100	40	100	60	100
VAB-nAMD-Physician	80	100	0	17	60	100	20	100	40	100
VAB-ALL-Patient	100	100	0	43	100	100	100	100	50	100
VAB-DME-Patient	0	100	0	33	50	100	0	100	0	100
OCR-MS-Nurse	40	100	0	40	100	100	20	100	0	100

Table 6: Segment-level performance comparison. Metrics: Used = reused claims, MTV = variation usage, Ind = indication, Type = content-type, Aud = audience.

Figure 4: Improvement rates from the rule-based baseline to the Golden Data Model across customer segments. We report gains in claim usage, MTV utilization, indication alignment, content-type alignment, and target-audience alignment. The most pronounced improvements occur in audience and content alignment , suggesting that structured contextual retrieval better supports personalized claim generation.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Provider/Model	Cost / 1K tokens
Gemini 2.5 Pro	1.41¢– 1.49¢
Gemini 2.5 Flash	0.21¢– 0.34¢
Anthropic (Claude)	0.32¢– 0.37¢

Appendix

Table 8: Example per-request costs from system logs. Costs depend on provider/model tier, motivating cost-aware routing in multi-pass verification.

Dimension	Evaluation Method	Range
Clarity	LLM, 7 weighted heuristics	$[0,1]$
Completeness	LLM, 5 weighted heuristics	$[0,1]$
Consistency	LLM, 4 weighted heuristics	$[0,1]$
Connectivity	Data-driven linkage checks	$[0,1]$
Compliance	Lifecycle and connectivity checks	$[0,1]$
Context Availability	Metadata availability checks	$[0,1]$

Appendix

Table 9: Overview of the 6C scoring dimensions. Each dimension is normalized to

[0,1]

and combined into the final 6C score.

Dimension	Type	Heuristic / Criterion	Weight	Interpretation
Clarity	LLM	Structure	0.15	Clear subject, verb, and outcome
		Quantification precision	0.15	Precise numerical or statistical values
		Clinical specificity	0.15	Specific endpoints, adverse events, or targets
		Context completeness	0.15	Population, setting, or usage context is defined
		Actionability	0.15	Clinically or practically usable
		Scientific alignment	0.15	Aligned with reference evidence

Appendix

Table 10: Unified 6C scoring rubric. Clarity, Completeness, and Consistency are evaluated with LLM-based weighted heuristics, while Connectivity, Compliance, and Context Availability are computed from graph, lifecycle, and metadata signals.

S_{6C}(x_{i})=\frac{1}{6}\left(s_{i}^{\mathrm{clar}}+s_{i}^{\mathrm{comp}}+s_{i}^{\mathrm{cons}}+s_{i}^{\mathrm{conn}}+s_{i}^{\mathrm{compl}}+s_{i}^{\mathrm{ctx}}\right)

Appendix

Algorithm 1 6C Claim Quality Scoring

6C Lens	Observed Pain Point	Example
Clarity	Frequent use of clinical shorthand and noun-only claims, which limits standalone interpretability.	CL-US-000853 : “T1 Gd+ lesions” lacks a verb, benefit, or complete claim structure.
Connectivity	Claims are often linked to materials, but parent-child claim relationships remain weak, making it difficult to verify whether semantically similar claims communicate the same message.	137 unique materials were tracked to claim IDs; 236 of 480 claims were referenced in at least one material.
Context	Statistical or numerical claims may appear without required qualifiers or supporting context, increasing the risk of unsupported downstream generation.	CL-US-000132 : numerical claims may “float” without the supporting messaging required for CMLR review.
Consistency	Overlap across formulations or variants can create identity conflicts when similar wording is reused across distinct product contexts.	Identical VA access claims were observed for IV and SC formulations.
Completeness	Long-form claims, especially safety-related claims, are prone to truncation and loss of legally relevant details.	CL-US-004992 : detailed IgG monitoring instructions are vulnerable to truncation.
Compliance	Core efficacy claims generally show strong reference-to-use linkage, but compliance still depends on lifecycle status and evidence availability.	High reference count was observed for core efficacy statements.

Appendix

Table 11: Failure mode matrix for Veeva claim records. The table summarizes observed data quality and governance pain points that motivate the Golden Data Model.

Dimension	Definition	Core Question	What It Evaluates
Clarity	Linguistic clarity and interpretability	Can I understand this?	Precision, readability
Completeness	Self-contained structured claim	Is everything included?	Structure, entities, outcomes
Consistency	Alignment across claim variants	Is it consistent across versions?	Message fidelity
Connectivity	Strength of supporting evidence	How strong is the evidence?	Reference linkage
Compliance	Regulatory validity	Is it allowed to be used?	Approval + evidence
Context	Availability of metadata	Is required info present?	Audience, material, metadata

Appendix

Table 12: Definition of the 6C evaluation framework.

Figure 5: Rule-heavy Composer setup before the Golden Data Model. The LLM receives separate prompt policies, citation constraints, and static claim rule book. Although this design enforces basic compliance, it does not provide structured audience, material, usage, or claim-quality signals, leading to generic outputs, audience mismatch, weak personalization.

Field	Value
Claim ID	CL-US-002291
Create Date	2023-10-17T19:48:36.000Z
Modified Date	2024-08-10T14:14:09.000Z
Product	Ocrevus
Country	United States
Match Text (Original Claim)	OCREVUS is proven effective in reducing relapses in RMS, slowing disability progression in RMS and PPMS, and reducing brain lesions in RMS and PPMS

Appendix

Table 13: Example claim record (CL-US-002291) represented in the Golden Data Model schema format.

Figure 6: Comprehensive Graph View of Claims.

System Prompt	Content
Role and Context	Process claim-level data into a structured data model. Extract information from each claim match text to populate layer_ai_guardrails_hallucination_prevention . Derive claim_specific_business_rule by combining the reference text and match claim text. Populate marketing_vector_store_content with tone signals such as patient-friendly, benefit-led, or HCP-focused. The primary_message_type , such as Efficacy, should be inherited from the claim classification. Populate prohibited_words_negative_constraints by extracting important phrases and phrases to avoid or not overstate from the original claim text.
Output Format	Output should be an Excel file. Keep the output structured at the claim-row level, not one row per unique claim ID. Include the following fields: layer_ai_guardrails_hallucination_prevention , claim_specific_business_rule , marketing_vector_store_content , and prohibited_words_negative_constraints .
Tool and Processing Policy	Do not rewrite the claims themselves. Match each row by its own claim ID and target audience combination. Do not collapse multiple rows into one row per claim ID, since the true unique unit is the combination of claim ID and target audience.
AI Guardrails Hallucination Prevention	{‘‘rule_type’’: ‘‘evidence_alignment’’, ‘‘derived_from’’: [‘‘Reference Texts’’, ‘‘Match Text’’], ‘‘rule’’: ‘‘Keep wording aligned to the match text and supported by the available reference text.’’}
Alignment	{‘‘rule_type’’: ‘‘audience_alignment’’, ‘‘target_audience’’: ‘‘Patient/Consumer-facing’’, ‘‘rule’’: ‘‘Prefer plain-language wording and avoid adding HCP-only terminology unless already present in the source text.’’}
	{‘‘rule_type’’: ‘‘material_alignment’’, ‘‘material_type’’: ‘‘Post Service / Mail’’, ‘‘rule’’: ‘‘Keep the wording appropriate for the material type without changing claim meaning.’’}

Appendix

Table 14: Foundation Prompt System Instruction.

Panel of Figure 8 — Figure 8: Post-deployment Composer usage patterns. The analyzed logs contain 16,963 requests, 146 million tokens, and an estimated monthly API cost of $543.39. The temporal distribution suggests human usage.

Personalized Text Generation

Personalized text generation adapts language-model outputs to a user's writing style, preferences, history, and context.

See all 22 papers

Jul 6, 2026

Medi-Gemma: A Hybrid Clinical Decision Support System Integrating Deterministic EMR Analytics and Retrieval-Augmented Generation

Deploying Large Language Models (LLMs) in high-stakes clinical settings remains limited by structural hallucinations, weak deterministic reasoning over tabular patient data, and omissions in vector retrieval. This paper presents the architecture and validation of Medi-Gemma, a Clinical Decision Support System (CDSS) for wound pathology triage and workflow automation. The platform introduces a decoupled framework that separates clinical perception from data orchestration while preserving traceable reasoning. Medi-Gemma uses a multi-stage pipeline coordinated by a centralized ClinicalOrchestrator. Data requests are handled without generative inference by a DataManager that cleans unstructured Electronic Medical Record (EMR) files through type coercion. Natural language queries are processed by a hierarchical IntentRouter, which routes requests to deterministic analytics paths executed by a PandasQueryEngine or to patient-specific reasoning managed by a ClinicalRAGEngine using a CPU-optimized vector store. A key contribution is the Ground Truth Injection Module, which intercepts patient-specific queries, extracts numeric identification tokens, queries the structured dataframe via Pandas, retrieves the latest validated clinical state, and embeds this snapshot as an overriding context block in the LLM prompt before generation. Safety compliance is enforced by a deterministic ProtocolManager that maps clinical terminology to fixed evidence-based risk pathways, while a SafetyVerifier phrase filter prevents output rule violations. Validation shows that this architecture eliminates semantic context drift, prevents database compilation crashes, and improves factual adherence to backend clinical repositories. These results support Medi-Gemma as a safer pattern for LLM-based clinical decision support where structured data fidelity, retrieval grounding, and deterministic safeguards are essential.

Mohammed Saim Ahmed Quadri, Yunzhe Xue, Justin W. Ady +1

Department of Computer Science, New Jersey Institute of Technology Newark, NJ, USA · Department of Data Science, New Jersey Institute of Technology Newark, NJ, USA · Vascular and Endovascular Surgery, Robert Wood Johnson Hospital New Brunswick, NJ, USA

Retrieval-Augmented Generation LLM Hallucination Mitigation Clinical Decision Support Clinical Reasoning Medical LLMs

Ask PDF

May 31, 2026

Benchmarking Local LLMs for Natural-Language-to-SQL Querying in Biopharmaceutical Manufacturing: An Empirical Benchmark on Consumer-Grade Hardware

Biopharmaceutical manufacturing organizations operate under regulatory frameworks such as FDA guidance, EU Good Manufacturing Practice (GMP), and the EU AI Act, which can restrict the use of cloud-based artificial intelligence systems. Locally deployed large language models (LLMs) offer a privacy-preserving alternative, but their suitability for pharmaceutical manufacturing tasks remains underexplored. This study evaluates four open-source LLMs (Qwen 2.5 Coder 7B, Llama 3.1 8B, Mistral 7B, and Meditron 7B) deployed locally via Ollama for natural-language-to-SQL generation over a pharmaceutical manufacturing database. A FastAPI-based evaluation platform, PharmaBatchDB AI, was developed using a synthetic Microsoft SQL Server database containing approximately 63,000 records across Batch, Manufacturing Execution System (MES), and Clean-In-Place (CIP) modules. Models were benchmarked on 60 domain-specific natural-language questions using metrics including SQL extraction rate, SQL compliance, factual consistency, ROUGE-L, hallucination rate, throughput, and latency. Qwen 2.5 Coder 7B, Llama 3.1 8B, and Mistral 7B generated SQL for all evaluation tasks, while Meditron 7B failed on nearly all tasks due to context-window limitations and poor SQL generation capability. Llama 3.1 8B achieved the highest SQL compliance, whereas Qwen 2.5 Coder 7B achieved the strongest overall text similarity and factual consistency. Performance differences between the two leading models were not statistically significant. The results show that code-tuned general-purpose LLMs outperform a domain-specific biomedical model on structured query generation for pharmaceutical manufacturing data. Although fully local, GxP-aligned NLQ systems are feasible on consumer hardware, current performance levels still require human oversight and downstream validation for regulated use.

Sagar Bhetwal, Rajan Bastakoti, Nirajan Acharya +2

Department of Computer Science, University of the Cumberlands, Kentucky, United States · Department of Computer Science, DePaul University, Chicago, IL, United States · Youngstown State University, Youngstown, Ohio, United States +1

LLM Evaluation LLM Inference Text-to-SQL

Ask PDF

Feb 12, 2026

Are Two LLMs Better Than One? A Student-Teacher Dual-Head LLMs Architecture for Pharmaceutical Content Optimization

Large Language Models (LLMs) are increasingly used for content creation in regulated domains such as pharmaceuticals, where content must be medical accurate and legally compliant. Traditional manual quality control (QC) is slow, error-prone, and create publication bottlenecks. To address this, we introduce a modular, LLM/VLMs-driven QC architecture, LRBTC (Language, Regulatory, Brand, Technical, Content Structure), implemented through a Student Teacher HITL architecture and Waterfall rule filtering logic for scalable, verifiable content validation. This architecture is designed to provide verifiable, traceable, and scalable content optimization. Our methods achieve 83.0% F1 and 97.5% recall on AIReg Bench, reducing missed violations five fold compared with Gemini 2.5 Pro. On CSpelling, it improves mean accuracy by +26.7 %. Our error analysis further reveals that while current models are strong at detecting misspellings (92.5% recall), they fail to identify complex grammatical (25.0% recall) and punctuation (41.7% recall) errors, highlighting a key area for future work. This work provides a practical, plug-and-play solution for reliable, transparent quality control of content in high-stakes, compliance-critical industries. We also provide demo access.

Suyash Mishra, Qiang Li, Anubhav Girdhar

Roche · Accenture · Involead

AI Regulatory Compliance LLM Reliability Language Model Error Detection

Ask PDF

Abstract

Personalized Text Generation

Explore similar work

Medi-Gemma: A Hybrid Clinical Decision Support System Integrating Deterministic EMR Analytics and Retrieval-Augmented Generation

Benchmarking Local LLMs for Natural-Language-to-SQL Querying in Biopharmaceutical Manufacturing: An Empirical Benchmark on Consumer-Grade Hardware

Are Two LLMs Better Than One? A Student-Teacher Dual-Head LLMs Architecture for Pharmaceutical Content Optimization