Financial Due Diligence RAG Prompt

This prompt is designed for M&A analysts, investment due diligence teams, and financial auditors who need to generate structured findings from a corpus of data-room documents. The core job-to-be-done is transforming raw, retrieved financial text—spanning contracts, P&L statements, board minutes, and audit letters—into a diligence report that surfaces red flags, identifies financial trends, and anchors every claim to a specific source document. The ideal user is a financial professional who understands the deal context and can validate the AI's output, not a general consumer asking open-ended questions about a company.

You should use this prompt when you have a pre-retrieved set of document chunks from a vector database or search index, and you need a synthesis that goes beyond simple Q&A. It is appropriate when the output must include explicit document references, distinguish between historical facts and forward-looking statements, and flag missing information that prevents a complete answer. The prompt is built for high-stakes workflows where an unsupported claim can have material consequences. It expects the input context to contain financial data, dates, and entity names that can be cross-referenced. Do not use this prompt for general market commentary, real-time news analysis, or tasks where the source documents are not provided in the prompt context.

This prompt is not a replacement for a financial model or a human diligence lead. It will not perform quantitative analysis, calculate IRR, or build a three-statement model. Its strength is in reading, summarizing, and pattern-matching across unstructured text. Before deploying, you must pair it with a robust retrieval pipeline that surfaces relevant chunks, and you must implement a human review gate for any finding that is flagged as a red flag or material risk. The next section provides the copy-ready template you can adapt to your specific data-room structure and output schema.

The following prompt template is designed to be wired into a Retrieval-Augmented Generation (RAG) pipeline for financial due diligence. It instructs the model to act as a senior financial analyst, synthesizing findings exclusively from the provided document chunks. The template uses square-bracket placeholders for all dynamic components—retrieved context, user questions, output schemas, and risk-level controls—so it can be adapted to different deal types, jurisdictions, and data-room configurations without rewriting the core instructions.

textCopy
You are a senior financial due diligence analyst reviewing data-room documents for an M&A transaction. Your task is to answer the user's question using ONLY the provided document excerpts. You must not use any external knowledge, training data, or assumptions beyond what is explicitly stated in the context.

## CONTEXT
[CONTEXT]

## USER QUESTION
[QUESTION]

## INSTRUCTIONS
1. Answer the question using only information present in the CONTEXT above.
2. For every factual claim, provide a document reference in the format [Doc: <document_name>, Section: <section_reference>, Date: <document_date>].
3. If the context contains financial figures, include the exact figure and the date it was reported.
4. Identify any red flags, inconsistencies, or unusual trends present in the context, and cite the specific documents where they appear.
5. If the context is insufficient to answer the question fully, explicitly state what is missing and what additional documents would be needed.
6. Distinguish between historical financial data and forward-looking statements or projections.
7. Do not provide investment advice, valuation opinions, or deal recommendations.

## OUTPUT FORMAT
[OUTPUT_SCHEMA]

## CONSTRAINTS
- Maximum of [MAX_FINDINGS] findings per response.
- Direct quotes from documents must not exceed [MAX_QUOTE_LENGTH] words.
- If confidence in any finding is below [CONFIDENCE_THRESHOLD], flag it with [LOW_CONFIDENCE] and explain why.
- Risk level for this analysis: [RISK_LEVEL]. Adjust the depth of scrutiny and caveat language accordingly.

To adapt this template for your pipeline, replace each square-bracket placeholder with values from your application layer. The [CONTEXT] placeholder should receive the top-k retrieved chunks from your vector store or search index, formatted with document metadata (name, section, date) prepended to each chunk. The [OUTPUT_SCHEMA] placeholder should contain a structured format specification—typically a JSON schema with fields for findings, document_references, red_flags, missing_information, and confidence_assessments. For high-stakes deals, set [RISK_LEVEL] to "high" to trigger more conservative language and explicit uncertainty flags. Before deploying, validate that your retrieval pipeline is attaching accurate metadata to each chunk; the prompt's citation discipline depends on it. Run a set of golden-test questions against a known document set and check that every cited reference resolves to the correct source location. If your eval shows citation errors exceeding 5% of claims, revisit chunk attribution before tuning the prompt further.

This prompt is designed to operate inside a RAG pipeline where a retrieval step has already returned a set of document chunks from a data-room index. The application layer is responsible for assembling the prompt with the correct inputs: the user's question, the retrieved context, and any output schema constraints. Before calling the model, the harness should validate that the retrieved context is non-empty and that the combined prompt length does not exceed the model's context window. If the retrieval step returns zero results, the system should surface an evidence-gap response rather than passing an empty context block to the model, which would force the model to either fabricate or refuse without useful diagnostic information.

The output from this prompt must pass through a structured validation layer before it reaches the user. Parse the model response as JSON and validate that every finding object contains a non-empty document_reference field that maps to a document ID present in the retrieved context set. Reject any finding where the red_flag boolean is true but the evidence_quote field is empty or truncated. For numerical claims in the trend_observation field, run a secondary check: extract any percentage or currency values and verify they appear in the source chunks. A mismatch does not automatically reject the finding, but it should downgrade its confidence score and add a numerical_verification: pending flag for human review. Log every validation failure with the raw model output, the specific field that failed, and the document IDs that were available at generation time.

Model choice matters for this workflow. The prompt requires strict JSON adherence, long-context reasoning over financial documents, and disciplined refusal when evidence is insufficient. Prefer models with strong instruction-following benchmarks on structured output tasks. If using a model that does not support native JSON mode, wrap the call in a retry loop with a repair prompt that feeds validation errors back to the model for correction. Set a maximum of two repair attempts before escalating to a human review queue. For high-stakes diligence workflows—such as pre-acquisition red-flag reports or material weakness identification—always route outputs with red_flag: true or confidence: low to a human analyst before the finding appears in any downstream system. The harness should also log the full prompt, retrieved document IDs, model response, validation results, and human review decision for audit trail purposes. This traceability is essential when diligence findings are later challenged or when the system's behavior needs to be reproduced for regulatory or internal review.