RAG Development · New York, NY

Custom RAG development in New York.

A first-year associate at a Big Law firm in Midtown bills out at $400 to $500 an hour. A credit analyst at a mid-size asset manager covers 300 issuers, each with a 10-K, four 10-Qs, six to ten 8-Ks, an annual proxy, earnings transcripts, and a stack of sell-side research. Both are spending the bulk of their week on document search.

That isn't a headcount problem. It's a retrieval problem. We build RAG systems for New York finance, legal, and healthcare firms that answer questions over your own document library in under ten seconds, with citations that point to the paragraph, not the document.

Engagement scope and pricing depend on document volume, source count, and the accuracy bar your compliance team requires.

How we work

Tell us which document library you need to query.

Where the hours actually go.

A credit research team at a New York asset manager tracks hundreds of issuers. The standard workflow looks like this: an analyst opens EDGAR, pulls the latest 10-K, ctrl-F's for "debt maturity", opens the deferred income tax footnote, opens last quarter's 10-Q to compare, opens the most recent earnings transcript for management commentary, and cross-references an internal credit memo from 2022. Twenty minutes per question, repeated forty times a day across the team.

In Big Law the cost is more visible because it's billed. Reviewing 200 commercial lease abstracts for a single private equity diligence project to surface every co-tenancy clause referencing Macy's or Bloomingdale's as anchor tenant is a $40,000 line item if it goes to associates. The work is low-judgement, high-recall pattern matching, which is exactly what a retrieval pipeline is built for.

Healthcare systems run into the same wall. Mount Sinai, NYU Langone, and Northwell each maintain enormous libraries of clinical policies, payer contracts, and compliance bulletins. A revenue cycle manager asking "does our 2024 Aetna contract permit out-of-network billing for DRG 470" is doing the same retrieval task as a JPMorgan analyst looking up a covenant, with the added constraint that the answer has to be auditable.

Generic ChatGPT-style search can't do this. It has no access to your documents. Vendor portals from Bloomberg or Westlaw answer general questions but don't know your internal memos, your contracts, or your historical positions. A RAG system sits on top of your private corpus and returns the exact paragraph that supports the answer.

What we build for New York firms.

Six concrete components, each chosen for the document types common in financial services, law, and clinical operations.

SEC EDGAR ingestion pipeline

Direct EDGAR API integration with automatic backfill across 10-K, 10-Q, 8-K, DEF 14A, and S-1 filings. Filings are parsed with XBRL-aware extraction so structured financial fields (revenue, EBITDA, debt schedule) stay queryable as numbers, not strings.

Table-aware document parsing

pdfplumber for native PDFs, unstructured.io for scanned or messy filings, and a custom layout model for footnote tables. Row, column, and section context is preserved so a question about Q3 net interest margin returns the correct cell.

Hierarchical chunking for contracts

Commercial agreements have article and section structure. We chunk at the clause level while carrying parent-section metadata, so a question about indemnification limits returns the exact clause with its governing article and definition cross-references.

Hybrid retrieval with metadata pre-filters

BM25 keyword search combined with dense vectors (text-embedding-3-large by default, Voyage AI for finance-specific fine-tuning). Metadata filters on issuer ticker, filing date, document type, and jurisdiction run before the vector search, keeping precision high without sacrificing recall.

Re-ranker and citation builder

Cohere Rerank or BGE Reranker on the top-50 retrieval set raises top-5 precision by 10 to 15 percent on financial and legal corpora. Citations render with document name, filing date, page, and paragraph index for direct copy into compliance memos.

Evaluation harness and audit log

Before handoff we build a test set of 100 to 200 real questions from your analysts and report recall at 5, MRR, and answer accuracy. Every production query is logged with the chunks retrieved and answer returned to satisfy NY DFS Part 500 and FINRA retention requirements.

Built for New York document workflows.

New York is a document town. Goldman Sachs, JPMorgan, BlackRock, and Bloomberg run on filings, research, contracts, and memos. Skadden and Cravath bill hours against document review. Palantir and Squarespace deal with enterprise contracts. Mount Sinai and NYU Langone navigate payer agreements and clinical policies. The common thread is that the answer is always sitting in a PDF somewhere; the bottleneck is finding it.

The regulatory overlay matters here in ways it doesn't everywhere. NY DFS Part 500 imposes cybersecurity controls on any system touching nonpublic financial information. FINRA Rule 4511 requires six-year retention on books and records that an AI system effectively creates when it surfaces an answer. HIPAA kicks in for any clinical RAG deployment. We design for these constraints from the first architecture sketch, deploying inside your VPC with KMS-managed encryption, role-based access at the chunk level, and full query audit logs.

Commercial real estate is the quiet third vertical. CBRE, JLL, Cushman, and the in-house teams at every major REIT spend meaningful hours on lease abstracts, estoppel certificates, and SNDAs. A retrieval system over a 2,000-lease portfolio pays for itself the first time a partner asks "which leases have a going-dark clause tied to anchor cotenancy" and gets an answer in seconds instead of three associates and a weekend.

Questions from New York buyers.

Can the RAG system actually answer questions over a 10-K including footnoted tables?: Yes, but only with a non-default extraction path. The financial tables inside a 10-K (segment revenue, deferred tax reconciliations, lease maturity schedules) are the exact cells your analysts care about, and standard PDF loaders flatten them into garbled text. We ingest filings using table-aware parsers (pdfplumber for clean PDFs, unstructured.io for messier ones, and direct XBRL parsing for structured fields). Each table cell is preserved with its row label, column header, and parent section heading. A query like 'what was Goldman Sachs net interest income in Q3 2024' returns the cell value plus a citation to the exact exhibit and table.
How do you handle the regulatory citation requirements that NY DFS and FINRA care about?: Every retrieved chunk carries metadata for source document name, filing date, section heading, page number, and paragraph index. When the system answers a question, the citation block in the response renders that metadata in a format your compliance team can paste directly into a memo or audit response. For FINRA-regulated workflows we also persist a query log (user, timestamp, question, chunks retrieved, answer returned) for the standard six-year retention period. NY DFS Part 500 cybersecurity rules apply to any system handling nonpublic information, so we deploy inside your VPC with audit logging from day one rather than as an afterthought.
What does the system do when the answer isn't in the corpus?: It says so. Confidence-thresholded retrieval is the difference between a useful internal tool and a liability. We tune a minimum similarity score per use case (typically 0.72 to 0.78 on cosine similarity with text-embedding-3-large) and combine that with a re-ranker score from Cohere Rerank or BGE. If neither pass the threshold, the system returns 'no high-confidence answer in the document library' along with the closest near-misses. For a credit analyst at JPMorgan, a hallucinated covenant clause is worse than no answer at all.
How long does a New York engagement typically take from kickoff to a system real users can query?: Most engagements run six to ten weeks from signed agreement to production. The first two weeks are document ingestion and chunking (a credit research corpus of 50,000 PDFs takes about that long with proper extraction). The next two to three weeks are retrieval tuning and the evaluation harness, where we build a test set of 100 to 200 real questions from your analysts and benchmark recall and citation accuracy. The final phase covers UI integration, access controls, and audit logging. Big Law engagements with privileged documents add roughly one week for matter-level access partitioning.

Ready to scope your pipeline?

Tell us what document types you have, roughly how many, and what questions your analysts or associates need to answer. We reply within one business day with a rough scope and price range.

Custom RAG development in New York.

Engagement scope and pricing depend on document volume, source count, and the accuracy bar your compliance team requires.

Where the hours actually go.

What we build for New York firms.

Six concrete components, each chosen for the document types common in financial services, law, and clinical operations.

SEC EDGAR ingestion pipeline

Table-aware document parsing

Hierarchical chunking for contracts

Hybrid retrieval with metadata pre-filters

Re-ranker and citation builder

Evaluation harness and audit log

Built for New York document workflows.

Questions from New York buyers.

Can the RAG system actually answer questions over a 10-K including footnoted tables?

Yes, but only with a non-default extraction path. The financial tables inside a 10-K (segment revenue, deferred tax reconciliations, lease maturity schedules) are the exact cells your analysts care about, and standard PDF loaders flatten them into garbled text. We ingest filings using table-aware parsers (pdfplumber for clean PDFs, unstructured.io for messier ones, and direct XBRL parsing for structured fields). Each table cell is preserved with its row label, column header, and parent section heading. A query like 'what was Goldman Sachs net interest income in Q3 2024' returns the cell value plus a citation to the exact exhibit and table.

How do you handle the regulatory citation requirements that NY DFS and FINRA care about?

Every retrieved chunk carries metadata for source document name, filing date, section heading, page number, and paragraph index. When the system answers a question, the citation block in the response renders that metadata in a format your compliance team can paste directly into a memo or audit response. For FINRA-regulated workflows we also persist a query log (user, timestamp, question, chunks retrieved, answer returned) for the standard six-year retention period. NY DFS Part 500 cybersecurity rules apply to any system handling nonpublic information, so we deploy inside your VPC with audit logging from day one rather than as an afterthought.

What does the system do when the answer isn't in the corpus?

It says so. Confidence-thresholded retrieval is the difference between a useful internal tool and a liability. We tune a minimum similarity score per use case (typically 0.72 to 0.78 on cosine similarity with text-embedding-3-large) and combine that with a re-ranker score from Cohere Rerank or BGE. If neither pass the threshold, the system returns 'no high-confidence answer in the document library' along with the closest near-misses. For a credit analyst at JPMorgan, a hallucinated covenant clause is worse than no answer at all.

How long does a New York engagement typically take from kickoff to a system real users can query?

Most engagements run six to ten weeks from signed agreement to production. The first two weeks are document ingestion and chunking (a credit research corpus of 50,000 PDFs takes about that long with proper extraction). The next two to three weeks are retrieval tuning and the evaluation harness, where we build a test set of 100 to 200 real questions from your analysts and benchmark recall and citation accuracy. The final phase covers UI integration, access controls, and audit logging. Big Law engagements with privileged documents add roughly one week for matter-level access partitioning.

Ready to scope your pipeline?

Tell us what document types you have, roughly how many, and what questions your analysts or associates need to answer. We reply within one business day with a rough scope and price range.