Skip to content
CodelessOps
Go back

The anatomy of a financial knowledge base

In July, Cerebras published a piece on the knowledge base they built for their engineers. It answers 15,000 questions a day. I read it twice, with two different hats on. The AI builder in me nodded along at the retrieval design, which is genuinely good. The accountant in me kept asking one question the article never has to answer: what happens when the answer goes in the board pack?

An engineer who gets a slightly stale answer loses an hour. A finance team that gets a slightly stale answer files it. The questions are different too. Not “where can I find X” but: what did we accrue for the licence renewal last year? Which entity carries the lease? What did we tell the board in March about cash runway? The person who knew has usually left, and the answer is split across a spreadsheet, a PDF and a policy nobody’s opened since the audit. The 2026 FP&A Trends survey puts the cost of this at 47% of the team’s time spent collecting and validating data. Nearly half the job is finding things.

So here’s the same kind of article, for a different species. The anatomy of a knowledge base built for finance: seven organs, what each one is for, and how I built each of them into Grounded. Then the part I enjoy most, what the animal can do once all seven are in place.

01 The corpus The packs, policies, filings and contracts. The record. 02 The notebook What the team decided and why. The judgment. 03 Ingestion Documents parsed, distilled, made searchable. Originals kept verbatim. 04 Retrieval Four search techniques, fused, then reranked. No single one is trusted. 05 Version sense FINAL beats draft, v3 beats v2, the reissue beats the original. 06 Verification Every figure tied back to the cell it lives in. The organ the others exist for. 07 Controls Golden questions, graded on demand, like any month-end control.
The anatomy. The first five organs exist in good engineering knowledge bases too. The last two are what make it financial.

Organ 1: the corpus

Cerebras spent most of their engineering on connectors. Slack ingestion over WebSockets, code repositories re-embedded on every commit, plugin scripts for team databases. They had to, because engineering knowledge lives in threads and diffs.

Finance is different, and the difference makes this organ simpler, not poorer. Finance knowledge is already document-shaped, on purpose. The pack is the record. The signed accounts are the record. Centuries of the profession have trained everyone to write the important things down in files, and those files are sitting in a drive right now: board packs, management accounts, statutory filings, accounting policies, forecasts and budgets, loan agreements, close checklists, prior-year workings, auditor correspondence.

So the corpus organ is upload, not integration. No connectors to babysit, no re-platforming, no behaviour change. What stays out matters just as much: the live ledger stays in the accounting system where it belongs, and chat stays in chat. If it moved this morning, it isn’t the record yet.

THE CORPUS · WHAT GOES IN Board_Pack_May.pdf Mgmt_Accts_May.xlsx Forecast_FY26_FINAL.xlsx Accounting_Policies.pdf Loan_Agreement_2024.pdf Close_Checklist.xlsx ONE QUERYABLE STORE Stays out: the live ledger, chat streams, anything still moving. The knowledge base holds the record, not the transaction system.
Every document lands in the same store and becomes queryable through the same interface. No connectors, no migration.

Organ 2: the notebook

Documents say what happened. They almost never say why.

“We treat the licence renewal as a prepayment because the contract runs to March.” “The auditor accepted this treatment in FY24, see their letter of 12 May.” “Watch the Q3 mix effect, it recurs every year the distribution segment outgrows the rest.” None of that lives in a filed document. It lives in the head of whoever did the work, which is exactly why the classic finance knowledge failure is a leaving date.

So notes are an organ, not a bolt-on. In Grounded they’re first-class citizens of search: embedded and retrieved alongside the documents themselves, so asking “how do we treat the licence renewal” surfaces the policy PDF and the analyst’s note about the exception, in the same answer. Team notes are shared with the team. Personal working notes stay personal. And when the person who wrote them moves on, the judgment stays queryable.

Organ 3: ingestion

This is the digestive tract, and most of it should be invisible. Documents get parsed (including OCR for the scanned PDFs every finance drive accumulates), then chunked along the document’s own structure. A board pack is headings, tables and narrative, so the splits follow headings, tables and narrative rather than cutting every thousand characters mid-sentence. Each document also gets distilled on the way in: an LLM extracts a title, a summary and topics, the same move Cerebras make with their Slack threads.

One rule here matters more than the clever parts. The original text is preserved verbatim, separately from anything the system adds for searchability. When an answer quotes a source later, it quotes the actual document, character for character, not a paraphrase of it. Remember that one. It’s the foundation organ 6 stands on.

Organ 4: retrieval

No single search technique survives contact with a real finance corpus. Ask “why did margin dip in Q3” and the answer in the pack says “adverse mix effect in the distribution segment.” Not one shared word. Only semantic search connects those two. Then the auditor pastes “IFRS 16 right-of-use” or a nominal code, and semantic search is suddenly the wrong tool, because a pasted code deserves an exact lexical match and nothing else should outrank it.

QUERY "Why did margin dip in Q3?" Board_Pack_Sep.pdf "Adverse mix effect in the distribution segment." MEANING MATCH No shared words. Wins. Meeting_Agenda_Oct.docx "Q3 margin: to be discussed at the next review." WORD MATCH ONLY Shares the words, answers nothing. 01 02 03 04 EXACT MATCH MEANING MATCH FUSION RERANK Pasted codes and standards win outright. Question and answer share no vocabulary. Every technique votes; consensus wins. A second model scores against the question.
Retrieval runs several views of the same corpus at once. Each technique covers a weakness of the others.

The organ works by refusing to pick a favourite. Lexical search and semantic search each produce their own ranked view of the corpus, the ranked lists get combined with reciprocal rank fusion (the same technique Cerebras describe, and I was pleased to find the same smoothing constant of 60 sitting in both codebases), and a reranking model then scores the merged candidates against the actual question. A chunk that merely shares vocabulary drops. A chunk that answers, rises.

Organ 5: version sense

Here the engineering playbook stops translating. Cerebras decay old answers by age, because a Slack answer about infrastructure from six months ago is probably describing something that no longer exists. Filings don’t work like that. A five-year-old lease is still the lease. Finance documents don’t expire, they get superseded, and the corpus fills up with the evidence: v2 sitting next to v3_FINAL, a DRAFT next to the issued pack, the April report and its reissue four days later.

ONE FIGURE · THREE FILES Forecast_Mar_v3_FINAL.xlsx ANSWERS COME FROM HERE Forecast_Mar_v2.xlsx SUPERSEDED · DEMOTED Forecast_Mar_DRAFT.xlsx DRAFT · DEMOTED
Version sense at work. Nothing is deleted or hidden; the authoritative file simply outranks its siblings.

So instead of age decay, this organ reads the way finance actually labels authority. Files marked draft, WIP, legacy or superseded get demoted. A lower-numbered version of the same file gets demoted when a higher one exists. The non-FINAL sibling of a FINAL file gets demoted. Ask about versions explicitly and the demotion switches itself off, because sometimes the superseded number is exactly what you’re after.

This isn’t theoretical for me. When I benchmarked my own tool on the Board Pack Test, one of the three fixes that took it from raw-model territory to full marks was precisely this: when two files carry the same figure, answer from the reissued FINAL, not the original it replaced.

Organ 6: verification

Cerebras’s pipeline ends at “answer + citations.” For engineers that’s enough. For finance it’s where the real work starts, because in finance a citation isn’t decoration. It’s the beginning of a tie-out.

Three behaviours make this an organ rather than a feature. First, citations quote the source verbatim and point at the place the figure actually lives, down to the sheet and cell in a workbook (this is why organ 3 kept the original text untouched). Second, there’s an independent check: press the button and a separate pass re-reads each citation against its source and grades it verified, partially supported, or contradicted. The same discipline you’d apply to a junior’s work before it goes in the pack, applied to the AI’s. Numbers get the extra treatment, traced to a source cell or to a calculation the system can show.

"Gross margin was 31.4% in September." Mgmt_Accts_Sep.xlsx · P&L!D18 VERIFIED 31.4% The cell the claim lives in. "Debtor days improved to 42." Mgmt_Accts_Sep.xlsx · whole file UNPROVEN A whole-file citation is not a proof. The checker sends it back.
The two states of a number. A claim tied to its exact cell can be confirmed by anyone; a claim pointed at a whole file can't.

Third, and hardest to build: abstention. When the documents don’t contain the answer to “what’s our Q4 covenant headroom”, the correct output is “your documents don’t say”, with no helpful nearby figures smuggled in. A knowledge base that guesses is a mild annoyance in engineering. In finance it’s a liability with a login page.

Organ 7: controls

How do you know any of the above keeps working after you add March’s documents? The same way finance knows anything keeps working: a control you can rerun.

The knowledge base carries a set of golden questions, each with an expected answer and the expected source documents. Run them on demand and a grader checks the figures, the citations and the refusals. It’s a reconciliation for the system itself, and it’s the organ I’d push hardest on if you’re evaluating anyone’s tool, including mine. I published my own version as the Board Pack Test: 34 documents, 25 questions, answer key and grader included. The frontier models got most answers right and banked 66 and 71 of the 81 machine-checkable points. The gap was never intelligence. It was proof.

What the body does

Anatomy is structure. The fun part is physiology: what a financial knowledge base can do once all seven organs are in place.

SKILLS The variance commentary, written the same way every month, from this month's files. A procedure that runs itself. DEEP RESEARCH The question that takes four board packs and an afternoon. Planned, delegated across the corpus, assembled with citations. ADVISORY Staged reviews with a defined output at each phase. Contract terms, covenant checks. Every claim still checkable.
Three things a living knowledge base does beyond answering questions. All three inherit the verification organ.

Skills turn a recurring workflow into a procedure the system runs the same way every time. Encode the variance commentary once, or the close checklist, or a covenant check, and it executes over the current month’s documents with the same citation discipline as any chat answer. Finance has always loved a documented procedure. This one performs itself.

Deep research handles the questions that aren’t one lookup. “Reconcile what the last four board packs said about cash runway against what actually happened” needs planning, several passes over the corpus, and an assembled answer with the evidence attached. That’s the difference between a search box and an analyst.

Advisory workflows run staged reviews, phase by phase, each phase producing a defined output. A contract review, an advisory pack. Long-form work, and the verification organ applies to all of it, not only to chat.

Here’s the part that ties the two halves together: none of the physiology is safe without the anatomy underneath it. A skill running on retrieval you can’t verify is automated guessing with a schedule.

Bring your own pack

Seven organs, then. Corpus, notebook, ingestion, retrieval, version sense, verification, controls. The first five make a good knowledge base. The last three (version sense, verification, controls) are what make it a financial one, and they’re the three the engineering playbooks skip, because engineering doesn’t file its answers.

Don’t take the anatomy on faith, mine or anyone’s. The Board Pack Test is public: documents, questions, answer key, grader. Better still, bring your own pack and ask the questions your team already knows the answers to. That’s a better exam than mine, because it’s yours.


Want answers like this from your own reports?I build finance teams a knowledge base of their own documents — every answer cited to the cell, accepted when it passes 25 questions your team wrote.

Share this post:

Keep reading

All posts →