How Fundry controls hallucinations

Every LP, CFO, and compliance officer who has evaluated an LLM product asks the same question: how do I know this thing won't make something up? This page is the concrete answer — the retrieval, citation, confidence, and approval controls that already run in production.

Read the full security whitepaper

The 60-second version

1Retrieval first. Every agent fetches your firm's real documents before it answers. The LLM is instructed to stay inside the retrieved passages.
2Citations are mandatory. Answers carry pointers back to the knowledge-entry rows they were derived from. No citations = review required.
3Confidence is measured, not claimed. Scores come from actual retrieval similarity, not model self-report. Low confidence gates the review queue.
4High-stakes output requires human approval. Compliance drafts, tax reviews, and DDQ responses all go through an approver gate before they leave Fundry.
5If we don't know, we say so. No-data responses are explicit and never quietly filled from the LLM's training corpus.

The controls in detail

Each of these is already operational. They are not a roadmap item — they are what the product does today.

Retrieval-grounded answers

Every Fundry agent answers from your firm's actual documents — not its training data.

Agents use a retrieval-augmented pipeline: before generating any answer, they fetch the most relevant passages from your knowledge base via vector search and exact-text lookup.
The model is explicitly instructed to answer only from those retrieved passages. If the passages do not cover the question, the agent says so instead of improvising.
Retrieved context is cited in the response, so every claim is attributable to a source.

Citation enforcement

Every structured answer includes citations back to the passages it was derived from.

DDQ responses, compliance suggestions, and tax-review outputs all carry citation objects pointing at the specific knowledge-entry rows that grounded the answer.
The frontend surfaces citations alongside each claim so reviewers can inspect the source before approval.
An answer with no citations is treated as low-confidence and routed through the review queue.

Confidence scoring

Every answer carries a numeric confidence score derived from retrieval similarity, not self-reported model certainty.

Confidence is computed from the vector-search similarity of the top retrieved chunks, plus whether the final answer actually references them.
Answers below a firm-configurable threshold (default 0.6) are flagged as review-required and not auto-applied to downstream artifacts.
Every confidence score is persisted so the product team can audit calibration over time.

No-data fallbacks

When Fundry doesn't know something, it says so — in writing.

If retrieval returns zero matching passages, the agent returns an explicit 'no data' response rather than filling the gap from the LLM's training data.
No-data responses are never counted as completed work, never contribute to knowledge-score milestones, and never close review queue items.
The UI clearly distinguishes 'no answer yet' from 'answer pending review' from 'approved'.

Human-in-the-loop on every high-stakes output

Compliance drafts, tax reviews, and DDQ responses require an approver before they leave Fundry.

Every compliance draft, Form D filing, blue-sky notice, and K-1 cover letter goes through an approve_reviews RBAC gate. A CFO or fund admin must click Approve before the artifact is marked final.
Approvals are persisted with the approving user id and timestamp. An approval is also a positive training signal — the approved artifact is written back to the knowledge base so future agents learn from it.
Suggestions and drafts can be edited in-place before approval. Edits are tracked against the original draft so the diff is auditable.

Provenance on every knowledge entry

Every fact in the knowledge base carries its source — document id, source type, confidence, approver.

The knowledge_entries table stores source_type ('firm_doc', 'system', 'manual'), source_ref (the document id), confidence, approved_by, approved_at, and the original agent that produced the row.
Ask Fundry's retriever reads these columns and only surfaces approved, high-confidence entries by default. Unapproved entries can be optionally included with clear labeling.
When a user asks Ask Fundry 'where did this come from?', the answer is a structured lineage, not a hand-wave.

What Fundry doesn't claim to do

Being useful is not the same as being magic. Here's an honest list of the places where Fundry can still get something wrong, and the guardrails that catch it.

Coverage gaps

Fundry only knows what you've taught it. If your firm's policy on a topic lives in an email thread that never made it into the knowledge base, Fundry won't be able to answer from it. We surface this as a 'no data' response, not a guess.

Cite, don't advise

Every compliance, tax, and legal output is labeled informational. We never represent Fundry as a substitute for licensed counsel — the disclaimer banner is non-dismissable on every compliance page for exactly this reason.

Stale knowledge

Regulatory rules, filing deadlines, and market conventions change. The system re-embeds and re-indexes on a rolling basis, but firm-specific facts (who your auditor is, what your standard K-1 cover language says) depend on you updating Fundry when they change.

LLM calibration drift

Foundation-model behavior can shift between provider versions. We pin model versions, track confidence calibration in an internal dashboard, and re-evaluate every major release before routing production traffic to it.

What you can verify yourself

We believe trust comes from inspection, not promises. Every one of the following is visible in the product today:

Citations on every DDQ answer, compliance suggestion, and tax review output — click any claim to see the source.
Confidence score visible on every generated artifact. Low scores trigger the review queue automatically.
Approval log on every compliance draft — who approved, when, what changed.
Knowledge-base browse view with source document, upload date, and confidence for every stored fact.
"Where did this come from?" button in Ask Fundry that returns structured lineage back to the original source.

Security architecture Try Fundry with your own documents