Earnings call transcripts are one of the most valuable datasets for investment research. Each quarter, thousands of public companies discuss their financial performance, strategic outlook, and market conditions. RAG transforms this unstructured data into a searchable, analyzable knowledge base that investment professionals can query in natural language.
Why Earnings Calls Are Ideal for RAG
Earnings call transcripts are lengthy, rich in quantitative and qualitative information, and follow a consistent structure (prepared remarks followed by Q&A). They contain forward-looking guidance, management sentiment, competitive positioning, and operational metrics that are critical for investment decisions. A single investment team covering 50 companies across 10 years would need to analyze 2,000 transcripts — a task ideally suited for RAG.
Transcript Sources
Earnings call transcripts are available from multiple sources: SEC EDGAR provides access to related filings, FactSet and Refinitiv offer comprehensive transcript databases with historical archives, Bloomberg Terminal provides integrated transcript access, and Financial Modeling Prep offers developer-friendly API access via its MCP server. The SEC EDGAR MCP server and Financial Modeling Prep MCP server provide direct access for RAG pipelines.
Optimal Chunking for Earnings Transcripts
Earnings call transcripts benefit from speaker-based chunking, where each speaker turn is a separate chunk. This preserves the attribution of financial commentary — queries like "What did the CFO say about margins?" retrieve exactly the CFO's comments. Use 512 tokens as the maximum chunk size for speaker turns, with metadata for speaker name, role, company, ticker, quarter, and year.
Metadata Strategy
Every earnings call chunk should carry rich metadata: company name and ticker, fiscal quarter and year, speaker name and role (CEO, CFO, analyst), section type (prepared remarks, Q&A), and conversation topic tags. This metadata enables filtered retrieval like "Show me only CFO commentary on gross margins from 2025 Q4 earnings calls."
Embedding for Earnings Text
Earnings call text has unique characteristics — it mixes financial terminology with conversational language, includes forward-looking statements, and contains industry-specific jargon. Voyage Finance-2 achieves the highest retrieval accuracy for earnings transcripts due to its financial domain training. OpenAI text-embedding-3-large is a strong alternative with broader ecosystem support.
Combining RAG with FinBERT for Sentiment
For earnings call sentiment analysis, combine RAG retrieval with FinBERT sentiment scoring. Retrieve relevant management commentary chunks using your RAG pipeline, then score each chunk's sentiment using FinBERT. This creates a powerful pipeline that identifies not just what management said, but how they said it — detecting subtle shifts in tone that might signal upcoming issues.
Example Query Patterns
- "What guidance did Apple provide for Q2 2026 revenue?"
- "Compare Microsoft's cloud revenue growth commentary across the last 4 quarters"
- "Show me all mentions of 'inflation' in JPMorgan earnings calls from 2024"
- "What did the CFO say about operating margins in the most recent quarter?"
- "Which companies mentioned AI investment in their Q1 2026 earnings calls?"