Key facts
| RAG status | Live - embeddings, retrieval and chat completions are available today |
| Embeddings | Embedding models are live for firm corpora and prior workpapers |
| Models | 30+ models behind one OpenAI-compatible API |
| Citations | Responses can cite retrieved source documents for review |
| Access control | Scoped API keys plus separated retrieval indexes per engagement |
| Deployment | Cloud, VPC, on-prem or air-gapped for confidential client data |
| Pricing model | Flat monthly self-serve plans; no per-token billing on self-serve |
| Endpoint roadmap | Files, batch and fine-tuning are coming soon |
TL;DR
- Index methodology and public tax guidance first, then extend to workpapers.
- Chunk by document structure, not a fixed character count.
- Require citations so reviewers can verify the passage behind an answer.
- Separate retrieval indexes per engagement to respect client confidentiality.
- Refresh the index when guidance changes or a new standard takes effect.
How it works, step by step
- Choose the first corpus: methodology manuals, tax guidance or engagement templates.
- Clean and chunk it by document structure, keeping headings and references intact.
- Generate embeddings and store them with source identifiers and effective dates.
- Build retrieval that returns passages plus citation metadata, not just text.
- Scope access per engagement so retrieval cannot cross client boundaries.
- Evaluate with real firm questions and require a preparer to review answers.
- Refresh the index on a schedule and record which document versions are live.
Try it yourself
Open the RAG architecture builder →
What RAG changes for accounting firms
Firms already hold the knowledge: methodology manuals, technical guidance, templates, and thousands of prior workpapers. RAG makes that corpus searchable in natural language and returns answers with source references, so juniors spend less time hunting through folders and more time applying judgment.
The practical gain is consistency. When every preparer retrieves from the same approved corpus, answers align with firm methodology instead of individual habit. Start with internal methodology and public guidance, where leakage risk is low, before connecting client workpapers.
Designing the knowledge base
Retrieval quality depends on chunking. Split documents by structure - section, standard, checklist item - so a retrieved passage is a coherent unit, and store metadata with each chunk: source document, section, effective date and confidentiality class. Hybrid retrieval, which combines keyword and vector search, handles reference numbers and defined terms well, a common need with tax and accounting material.
Keep the index current: superseded guidance should be replaced, not appended, so answers cite what is actually in force.
Access control per engagement
Client confidentiality shapes the architecture. Public guidance and firm methodology can share an index; client workpapers should sit in separate indexes with their own API keys, so a retrieval call for one engagement cannot surface another client's data. That mirrors the access model firms already apply to engagement files.
Log requests, retrieved source identifiers and reviewing staff so any answer that informed client work can be reconstructed. See AI audit logs for a schema, and use private deployment where client contracts require it.
Evaluation, refresh and review
Build an evaluation set from real firm questions with known correct answers and run it whenever the corpus, chunking or model changes. Require citations in every answer and have a qualified preparer verify before output informs client work. Refresh the index when standards change, and note that file upload and batch endpoints are coming soon, which will simplify bulk ingestion of document sets. The same OpenAI-compatible API means the pilot moves to production without code changes.
Honest comparison
| Capability | RAG over firm knowledge | Keyword search | Model without RAG |
|---|---|---|---|
| Answers | Drafts answers with cited passages | Returns document lists | Generates from training data |
| Currency | Refreshed index reflects current guidance | Depends on file naming | May use outdated knowledge |
| Verifiability | Citations tie answers to sources | Manual reading required | No source trail |
| Confidentiality | Index separation per engagement | Folder permissions | Corpus not involved |
| Effort | Setup and evaluation once, then reuse | Works immediately, limited recall | Fast but unreliable for specifics |
Frequently asked questions
What is RAG in an accounting context?
RAG - retrieval-augmented generation - retrieves relevant passages from your own documents and has a model draft an answer grounded in them, with citations. In an accounting firm the corpus is methodology, guidance and workpapers.
Can workpapers be used without exposing client data?
Keep client workpapers in separate indexes with their own scoped keys, and use private deployment where contracts require it. Retrieval then cannot cross engagements.
How should documents be chunked?
By structure - section, standard or checklist item - rather than a fixed character count, keeping headings and reference numbers with the text so retrieval and citations stay meaningful.
Do answers replace professional judgment?
No. RAG output is a draft with sources; a qualified preparer reviews it before it informs client work, and remains accountable for the conclusion.
How do we keep guidance current?
Refresh the index when standards or firm methodology change, replace superseded documents rather than appending them, and note the effective date in chunk metadata.
What does it cost to run?
Self-serve plans are flat monthly with no per-token billing, and the free plan includes plugsky-micro and plugsky-lite with no card. See the live pricing page for current plans.
Where should a firm start?
Pilot on internal methodology and public tax guidance with the free plan, evaluate with real firm questions, then extend to prior workpapers on private deployment.