How do you build finance assistants with RAG?

Key facts

Pipeline	Parse → chunk → embed → retrieve → cited answer
Embeddings	plugsky-embed and plugsky-embed-multilingual via /v1/embeddings (live)
Generation	Cited answers with JSON mode on /v1/chat/completions (live)
Long context	plugsky-longctx for comparing multiple periods in one prompt
Metadata	Entity, period, statement type and page stored per chunk
Models	30+ models behind one OpenAI-compatible endpoint
Residency	Region-locked planes, VPC, on-prem and air-gapped options
Free tier	plugsky-micro and plugsky-lite on the free plan, no card required

TL;DR

• Chunk filings by section and note, not by fixed length — finance meaning is structured.
• Store entity, period and statement metadata so retrieval can filter before ranking.
• Require citations to page and paragraph, and show the source text beside each figure.
• Never let the assistant compute totals; retrieve the reported number instead.
• Keep the corpus and prompts inside a region-locked plane for regulated data.

How it works, step by step

1. Collect the corpus — annual reports, accounting policy manuals, loan agreements, research notes — and license or classify each source.
2. Parse documents with layout awareness so tables, footnotes and headings survive; keep page numbers in metadata.
3. Chunk by section, note and table, storing entity, period, statement type and page alongside each vector.
4. Embed chunks with plugsky-embed (or plugsky-embed-multilingual for mixed-language corpora) and index them in your vector store.
5. At query time filter by entity and period, retrieve the top passages, and pass them to a chat model with an answer-only-from-context instruction.
6. Return a structured answer with a sources array pointing to document, page and section, and render sources next to the answer.
7. Evaluate on questions whose answers you can verify in the source documents, including questions the corpus cannot answer.

Original data

Try it yourself

Open the RAG architecture builder →

Why RAG fits finance documents

Finance questions are usually answerable from a specific paragraph, table or note — and the answer is worthless unless you can point to it. RAG matches that structure: retrieval finds the passage, generation explains it, and the citation makes it verifiable. Unlike a model's memory, the corpus can be versioned, audited and updated when filings change.

The model side runs on Plugsky's live endpoints: /v1/embeddings for indexing and /v1/chat/completions for grounded answers with JSON mode. Parsing, storage and filters stay in your infrastructure, which is what lets you enforce entity-level access and retention rules.

Chunking tables, notes and policies

Generic fixed-size chunking destroys finance structure. A revenue table split mid-row, or a note separated from its heading, produces confident nonsense. Chunk along document structure instead: one chunk per note, one per table with its caption, one per policy clause, with parent headings carried into metadata.

• Tables: convert to text or markdown with headers repeated, and keep the page number for citation.
• Footnotes: attach to the paragraph they qualify, or index separately with a pointer.
• Versioning: store fiscal period and filing date so superseded documents can be filtered or ranked lower.
• Exact identifiers: keep keyword search alongside vectors so account codes and clause references match reliably.

Citations, controls and evaluation

Ask for structured output: a direct answer, the retrieved values used, and a sources array with document, page and section. Instruct the assistant to say when sources conflict or when the corpus lacks the answer — silence is safer than a plausible guess. Render each figure with its source link so analysts can verify in one click.

Control access by entity and desk with pre-retrieval filters, and keep prompt retention aligned with your data policy. Because the model call carries document text, pin the workspace to a region-locked plane or deploy on-prem when rules require it. Evaluate with verifiable questions per statement type, score citation accuracy and refusal quality, and add documents to the golden set whenever a new corpus is indexed.

Honest comparison

Capability	Plugsky RAG	Keyword filing search	Model with no retrieval
Answer source	Retrieved passages with citations	Result list only	Model memory
Freshness	Re-index when documents change	Index-driven	Training cutoff
Traceability	Page and section per figure	Document-level	None
Multilingual	plugsky-embed-multilingual shares one vector space	Varies	Depends on model
Residency	Region-locked planes, VPC, on-prem, air-gapped	Your own systems	Vendor-controlled

Frequently asked questions

Can a RAG assistant calculate ratios and growth?

It should not calculate; it should retrieve reported figures and cite them. For derived metrics, compute in your own code and pass the result into the prompt as context.

How do you handle tables in PDFs?

Parse with layout awareness, convert tables to text with repeated headers, chunk per table with its caption, and store the page number so citations stay accurate.

Which embedding model suits filings?

plugsky-embed works for English-language filings; use plugsky-embed-multilingual when the corpus or queries mix languages such as Arabic and English.

How do we stop answers based on old filings?

Store fiscal period and filing date on every chunk, filter or demote superseded versions at retrieval, and re-index when a new filing replaces an old one.

Does Plugsky see our documents?

Only the passages you retrieve and send as prompt context. Documents and vectors stay in your storage, and prompt retention is configurable.

Can it run inside a bank's network?

Yes. Plugsky supports region-locked cloud planes plus VPC, on-prem and air-gapped deployments for regulated environments.

How should we measure quality?

Use verifiable questions per statement type and score citation accuracy, numeric fidelity to the source and refusals when the corpus lacks the answer.