Why Do Deterministic and Adaptive Retrieval Work for Credit Risk Management? (Part I)

1. Industry Context

Credit risk professionals operate in an environment where speed, precision, and defensibility are equally non-negotiable. Analysts, underwriters, and portfolio managers are expected to synthesize financial statements, ratings actions, management commentary, and market signals into judgments that carry real capital consequences. Increasingly, these professionals are turning to AI and large language models (LLMs) to accelerate research and analysis.

At the same time, credit analysis cannot tolerate hallucinated figures, stale data, or answers that cannot be traced to their sources. A single fabricated EBITDA figure or an outdated leverage ratio can materially skew a credit opinion. However, the breadth and complexity of credit analysis and risk management processes (spanning ratings, financials, transcripts, news, and market data) mean analysts often need to synthesize multiple sources quickly.

This dual reality is the central challenge that AI-assisted credit workflows look to solve. This is the reason S&P Global has architected two complementary retrieval approaches: Deterministic Retrieval and Adaptive Retrieval.

We focus on Deterministic Retrieval in this first instalment and take a deep dive into Adaptive Retrieval in our follow-up article.

2. Introducing Deterministic and Adaptive Retrieval

At the heart of this approach are two retrieval modes, each designed for a different kind of credit research question.

Deterministic Retrieval is built for specific, structured data requests. These are questions that have one correct, verifiable answer.

• Deterministic Retrieval maps the query directly to an API endpoint and underlying query against S&P Global's structured databases.
• Because the path from question to data is fixed and verifiable, every answer returned is sourced, reproducible, and token-efficient.
• This is critical for reducing ambiguity and cost in high-volume, structured analysis.
• An example prompt is: "What was Microsoft's EBITDA from 2021 through 2024?"

Adaptive Retrieval is designed for open-ended, research-driven queries that may require synthesizing information from multiple datasets at once.

• Rather than mapping directly to one data source, Adaptive Retrieval uses a router that directs the query to specialized retrieval agents to aggregate the data.  Each agent specializes in a particular dataset. The results are then assembled with citations, making Adaptive Retrieval well-suited to complex, multi-agent workflows and capable of serving as a single point of entry across the breadth of S&P Global's datasets.
• An example use case is combining a company's credit rating, management commentary from earnings transcripts, and broader market context from Market Intelligence News into a single, response.

Both retrieval modes are accessible through the S&P Global AI Data Portal, which can be integrated via a Model Context Protocol (MCP) server. This makes S&P Global's data (including the Kensho LLM-ready API) natively searchable in natural language by LLMs such as Claude and ChatGPT, as well as by agentic AI systems built on platforms like Microsoft Copilot, Amazon Quick Suite, Databricks, and Mistral.

3. How Do Deterministic and Adaptive Retrieval Work?

What makes retrieval deterministic is the fixed, predictable relationship between a query and its data source

• When an analyst asks a structured question (e.g., a metric, a company, a time range), the system understands the context based on use cases and inputs from domain experts managing individual datasets.
• The query maps directly to an API endpoint, which in turn executes a defined query against S&P Global's databases.
  • With well-specified queries (see prompting guide), deterministic retrieval provides consistent and reliable results. 

More importantly, the deterministic approach to data retrieval:

• Completely removes the possibility of hallucinations of ratings data in most cases.
• Is consistent and replicable, returning the same result each time the same question is asked.

These are attributes credit professionals need when reproducibility and audit trails matter, supporting use cases such as internal credit memos, investment tear sheets, or rating committee documents.

How Adaptive Retrieval can be used for credit analysis

However, credit research questions are frequently investigative, rather than limited to a single answer, e.g., "Build me a diligence snapshot on this issuer" or "What's driving the change in this company's credit profile?”. These questions cannot be resolved by a single API call; instead, the system needs to flexibly reach across ratings, commentary, and market data within a single workflow, adjusting its retrieval strategy to the question rather than requiring the analyst to know in advance exactly which dataset contains the answer. We will address this in our follow up article.

4. Case Study: Credit Ratings Data Retrieval

To illustrate how these two retrieval modes come together in practice, consider two example workflows referenced in S&P Global's own illustrations of the AI Data Portal in action.

Deterministic Retrieval Example on a Company’s Credit Ratings and Fundamentals

An analyst asks, " What were <insert company name> credit ratings and Debt/EBITDA ratios from 2021 through 2024?" (see Figure 1)

This structured request is mapped directly to different API endpoints, which execute a defined query against S&P Global's Ratings and financial databases. The result is a clean, sourced time series of EBITDA values tied back to the underlying filing or reporting records, along with credit ratings sourced directly from our databases.

The screenshots below show this query run live in an AI platform that is connected to the S&P Global AI Data Portal via the S&P Global connector. It shows a sequence of tool calls the model makes in response to the prompt: it searches for the relevant issuer credit ratings tool, retrieves its ratings history, then retrieves the financial line items needed to calculate Debt/EBITDA across the four fiscal years.

Figure 1a: Deterministic Retrieval Output featuring financials and S&P Global Credit Ratings

Figure 1 (cont’d): Deterministic Retrieval Output featuring financials and S&P Global Credit Ratings

5. How Is This Different from Generating Output Directly from LLMs Using This Data as a Corpus?

If ChatGPT, Claude, or Copilot can already ingest documents or be pointed at a data corpus, why not simply feed S&P Global data into these tools directly and let the LLM generate answers?

The differentiation lies in how the data is retrieved and represented in the final output, not merely whether the data is technically accessible to the model.

1. Traceability and auditability: When an LLM is simply given a large corpus and asked to answer a question, it synthesizes an answer probabilistically from patterns in the text it has ingested. Even when the underlying data is accurate, the model's generated answer may not point back to a specific, verifiable source. It may blend, paraphrase, or subtly misstate figures in the process (a hallucination risk that persists even with grounding data present).
   • Deterministic and Adaptive Retrieval provide transparency on the specific tools used to retrieve the records from S&P Global datasets. This makes the answer verifiable in a way that free-form LLM generation over a corpus is not.
2. Built-in governance and data quality controls: Retrieval agents are designed and curated by subject experts who understand how to extract the precise data points required for credit analysis. This includes handling inconsistent identifiers, missing metadata, and staleness issues that commonly degrade the reliability of naive LLM-corpus approaches. This built-in governance reduces the engineering burden that would otherwise fall on a firm trying to validate LLM outputs against a raw data dump.
3. Structured vs. probabilistic pathways (i.e. identical results across runs): Deterministic Retrieval's direct API-to-query mapping guarantees the same, reproducible answer to a structured question every time. An LLM generating an answer from a corpus may not produce identical wording or calculations every time, even when using the same underlying data. This introduces variability that is unacceptable in credit contexts where consistency and reproducibility are required for compliance and rating committee documentation.
4. Data is retrieved from S&P Global’s databases, not replicated from corpus of articles and text: The retrieval layer can be integrated into multiple GenAI applications (e.g., Claude, ChatGPT, Google Gemini, Microsoft Copilot, Quick Suite, Databricks, Mistral, etc.) as a live, directly retrievable service rather than a static, article text blob. This means the underlying data stays current and authoritative.

6. The Role of Data

Underpinning both retrieval modes is an understanding that regardless of the agentic AI models used or the sophistication of the platforms delivering the content – the quality, structure, and governance of the underlying data make or break the trustworthiness of the output. What makes data trustworthy for credit risk use cases?

• Structured, cited, and AI-ready data: S&P Global's datasets are structured and pre-cited specifically so that responses generated through the AI Data Portal are ready for professional use, supporting governance and compliance requirements rather than working against them.
• Domain experts involved in design and performance validation of retrieval agents: Each specialized retrieval agent used in Adaptive Retrieval is designed and curated by subject matter experts and engineers who understand both the dataset and real customer use cases. This ensures that the right data points (not merely the most textually similar ones) are extracted for credit analysis.
• Currency and authority: Because retrieval draws live data from S&P Global's databases rather than a static corpus that may become outdated, the information returned is current and authoritative. This directly addresses the "stale data" problem that plagues many LLM-based research tools.
• Addressing known failure modes: The architecture is explicitly built to counter the recurring challenges of AI in credit analysis: hallucinations, stale data, missing metadata, inconsistent identifiers, and limited auditability. By anchoring every response to structured, governed data with source citations, these failure modes are substantially mitigated rather than left to the LLM's discretion.
• Single point of entry across datasets: Rather than requiring analysts to know which of many S&P Global databases holds a given piece of information, the AI Data Portal (through Adaptive Retrieval's router) can serve as a unified entry point. This reduces the data-discovery burden on the end user.

7. Conclusion

Deterministic and Adaptive Retrieval represent two complementary answers to the same underlying problem: how do you bring the power of generative AI into a domain (credit risk management) that has a low tolerance for unverifiable, inconsistent, or stale information?

Deterministic Retrieval answers this for structured, well-defined questions by mapping queries directly to governed, verifiable data sources. Adaptive Retrieval answers this for open-ended, research-intensive questions by routing intelligently across specialized, subject expert-curated retrieval agents and returning fully cited, synthesized results.

Together, and often used in tandem within a single workflow, these two modes allow credit professionals to move from broad, cross-source research to precise, verifiable figures without leaving the AI environment they already work in, whether it is Claude, ChatGPT, Microsoft Copilot, or another integration via the Model Context Protocol.

This is fundamentally different from simply pointing an LLM at a data corpus: it is a purpose-built retrieval architecture, governed by domain expertise and designed to substantially mitigate hallucination risk while preserving the auditability that credit analysis requires.

As AI becomes further embedded in credit workflows across ratings, underwriting, and portfolio surveillance, the firms that succeed will be those that treat retrieval architecture (not just model capability) as the foundation of trust. Deterministic and Adaptive Retrieval, delivered through the S&P Global AI Data Portal, offer exactly that foundation: trusted credit intelligence, reliably brought into the AI workflows where analysts increasingly do their work.