Financial Report Q&A Prompt with Source Attribution

This prompt is designed for financial analysts, audit teams, and AI engineers building Q&A systems over financial filings, earnings reports, and disclosures. Its primary job is to force the model to answer strictly from provided context, attach exact section references, date-stamp all figures, and explicitly label forward-looking statements versus historical data. Use it when the cost of a hallucinated number or misattributed quote is higher than the cost of a non-answer—for example, when preparing audit evidence, drafting analyst notes, or generating investor communications from SEC filings.

The ideal user provides a complete [CONTEXT] block containing the full text of the relevant financial document sections, not just summaries or metadata. The prompt expects the model to refuse to answer if the context is insufficient, rather than inventing plausible-sounding figures. It is built for single-hop or multi-hop questions that can be resolved within the provided context, such as 'What was the Q3 2024 revenue for the Widgets segment?' or 'Identify all forward-looking statements regarding capital expenditure in the MD&A section.' The output contract requires structured JSON with fields for the answer, source citations including section names and paragraph references, date stamps for all numerical claims, and a boolean flag indicating whether the answer contains forward-looking information.

Do not use this prompt for general financial advice, real-time market commentary, or scenarios where retrieved context is not provided. It is not a replacement for a financial advisor, nor is it designed to interpret market sentiment or predict stock movements. The prompt explicitly instructs the model to abstain when context is missing, but it cannot detect whether the provided context itself is outdated or incomplete—that responsibility remains with the retrieval pipeline and the human operator. For production deployments, always pair this prompt with a retrieval system that fetches the correct filing version and date, and implement post-generation validation to verify that every cited figure can be traced back to the source text.

This prompt template is designed to be placed directly into your system instructions for a financial report Q&A application. It enforces strict source attribution, requiring every factual claim to be linked to a specific section, date, and filing. The template uses square-bracket placeholders that you must replace with your application's specific values before deployment. The core instruction set constrains the model to answer only from provided context, flag forward-looking statements, and explicitly state when information is unavailable.

codeCopy
You are a financial report Q&A assistant. Your answers must be precise, traceable, and grounded exclusively in the provided [CONTEXT].

## Core Rules
- Answer ONLY using information explicitly present in [CONTEXT]. If the answer is not in [CONTEXT], state "The provided documents do not contain this information." Do not guess or use external knowledge.
- For every factual claim, cite the source using the format: [Source: Document Name, Section, Date].
- Distinguish between historical data and forward-looking statements. Flag all forward-looking statements with: [FLS: This statement reflects management expectations as of [DATE] and is subject to risks and uncertainties.]
- If figures from different periods are compared, explicitly state the end dates for each period.
- If [CONTEXT] contains conflicting information, present both data points with their sources and flag the discrepancy.

## Input
[CONTEXT]:
"""
[RETRIEVED_DOCUMENTS]
"""

[QUERY]:
"""
[USER_QUESTION]
"""

## Output Format
Return a JSON object with the following schema:
{
  "answer": "string (the synthesized answer with inline citations)",
  "sources": [{"document": "string", "section": "string", "date": "string", "excerpt": "string"}],
  "forward_looking_statements": [{"statement": "string", "source": "string", "date": "string"}],
  "data_limitations": "string (any temporal, scope, or conflict caveats)",
  "abstention": "boolean (true if the answer could not be found in context)"
}

## Constraints
- Do not invent financial figures, dates, or ratios.
- Do not provide investment advice or interpretive commentary beyond what is directly stated.
- If a figure is preliminary, estimated, or restated, note this explicitly.
- Prioritize the most recent filing when temporal conflicts arise, but note the discrepancy.

To adapt this template, replace [RETRIEVED_DOCUMENTS] with your RAG pipeline's output, ensuring each chunk includes metadata for document name, section, and date. Replace [USER_QUESTION] with the end-user's query. Before production, validate that your retrieval system consistently provides the date-stamped metadata this prompt requires. Run a set of golden test queries through the assembled prompt and check that the output JSON strictly conforms to the schema, that citations are not hallucinated, and that forward-looking statements are correctly flagged. For high-stakes financial reporting workflows, always route outputs through a human review step before distribution.

The Financial Report Q&A prompt is designed to operate inside a retrieval-augmented generation (RAG) pipeline, not as a standalone chat interface. The application layer is responsible for document ingestion, chunking, retrieval, and post-generation validation before any answer reaches an analyst or audit team member. Treat this prompt as one component in a larger harness that includes a retriever, a numerical validator, a temporal consistency checker, and an audit logger.

Begin by assembling the prompt context from a retrieval step. For financial filings, prefer section-aware chunking that preserves the SEC item number, filing date, and document type in each chunk's metadata. When the user submits a question, run retrieval across the target filing set and include the top-k chunks with their metadata in the [CONTEXT] placeholder. If the question contains a temporal reference like 'Q3 2024,' inject a date filter into the retrieval query to avoid pulling stale figures. The [CONSTRAINTS] placeholder should enforce citation format (e.g., 'Item 7, FY2024 10-K, p. 42'), flag forward-looking statements, and require date-stamped figures. The [RISK_LEVEL] field should be set to 'high' for any question involving material financial figures, which triggers additional post-generation checks.

After the model returns an answer, run a structured validation layer before surfacing the output. Parse the answer for cited figures and cross-reference them against the source chunks using a numerical extraction validator. If a cited figure does not appear in the source text, flag the answer for human review and log the mismatch. Implement a temporal consistency check that verifies all date-stamped figures align with the filing period referenced in the question. For high-risk queries, route the answer to a review queue where an analyst can confirm or reject the response. Log every answer, its citations, validation results, and the reviewer's decision for audit trail purposes. Use a model with strong instruction-following and low hallucination rates on numerical data; Claude 3.5 Sonnet or GPT-4o are reasonable starting points, but run your own eval on a golden dataset of financial Q&A pairs before committing to a model.

Avoid wiring this prompt directly into a customer-facing chatbot without the validation harness. Financial figures carry material consequences, and even small hallucination rates on numbers can erode trust. If you are building for an internal audit team, integrate the prompt into a document Q&A interface that shows source passages alongside the answer so reviewers can spot-check citations. For SEC filing analysis, consider maintaining a filing-date index so the retriever never mixes figures from different reporting periods unless the question explicitly asks for a comparison. The next step after implementing the harness is to build a regression test suite using the eval criteria described in the Testing and QA section of this playbook.