← Writing & Research

Kuzushi Labs · Engineering notes

Engineering financial RAG

A hedge fund equity analyst platform, a million financial documents, and six weeks to deliver. Fourteen articles on the client platform—from customer requirements and parsing costs to evidence retrieval, paragraph citations and agentic research.

Arunank Sharan · October 2025 · 14 articles · 18,187 words

financial documents
1M+
parsing-cost reduction
60%
delivery window
6 weeks

The 2025 delivery brought a $120,000 parsing estimate to approximately $48,000 for the same corpus, within a $50,000 iteration budget. The series also develops current designs for bounded agents, evaluation and cloud operations.

Start with the analyst →

read or print the complete series · download the reading kit

  1. PART 01 · 7 min read

    A million documents, six weeks, and a $50,000 constraint

    Starting with the equity analyst’s workflow, then turning commercial constraints into engineering decisions.

  2. PART 02 · 7 min read

    A financial corpus begins with identity, time and provenance

    Sourcing filings and transcripts without losing amendments, availability dates, permissions or lineage.

  3. PART 03 · 7 min read

    A correctly read number can still be the wrong financial fact

    Tables, footnotes, units, reading order and the contract between a parser and a retrieval system.

  4. PART 04 · 8 min read

    Bringing a $120,000 parsing iteration inside a $50,000 budget

    The 2025 move from LlamaParse to Mistral, the 60% cost reduction, and what changes in 2026.

  5. PART 05 · 7 min read

    Chunk the evidence, not just the text

    Filing sections, speaker turns, table rows and the hidden difference between word and token budgets.

  6. PART 06 · 6 min read

    Choosing embeddings for financial evidence

    Financial BGE, SPLADE, Matryoshka dimensions and the cost of changing a million-document index.

  7. PART 07 · 6 min read

    Finding the right issuer, period and passage

    Dense and sparse retrieval, metadata constraints, rank fusion and candidate-depth experiments.

  8. PART 08 · 7 min read

    From fifty candidates to an answerable evidence set

    Cross-encoders, query rewriting, truncation, context packing and evidence diversity.

  9. PART 09 · 7 min read

    A citation should take the analyst to the evidence

    Paragraph identity, PDF coordinates, HTML highlighting and the interface that makes answers inspectable.

  10. PART 10 · 6 min read

    Make the calculation inspectable, too

    Grounded synthesis, financial units, deterministic arithmetic, contradictions and streaming answers.

  11. PART 11 · 7 min read

    When financial research needs more than one retrieval step

    A bounded agentic RAG design with typed tools, an evidence ledger, cost limits and explicit stopping rules.

  12. PART 12 · 7 min read

    Evaluate the evidence chain, not just the final answer

    Parser, retrieval, calculation, citation and agent evaluations that explain where quality was lost.

  13. PART 13 · 9 min read

    Operating RAG across a million financial documents

    Durable ingestion, separate serving paths, latency budgets, AWS and GCP deployment designs.

  14. PART 14 · 7 min read

    The engineering decision that made the product affordable

    Six-week delivery, parsing economics, analyst value, further funding and the next system improvements.

For the project overview, read the client case study. For the speech systems series, explore Engineering conversational AI.