Case study: private wealth intelligence
A private wealth copilot grounded in your own records.
Use RushDB as the evidence graph and vLLM as the local inference layer so family-office and VC teams can ask portfolio questions against holdings, deal memos, policies, and diligence notes without sending private context to a hosted LLM.
Private wealth teams cannot treat a chatbot as an investment system.
Family offices and VC teams work across portfolio records, cap tables, research notes, committee memos, policies, and private deal flow. A general chat interface can sound confident while missing the evidence, provenance, access boundary, or investment policy that should govern the answer.
Before
- Portfolio data, memos, and diligence notes live in separate systems
- Hosted model calls create data-residency and confidentiality questions
- Answers are hard to audit back to source records
- Unsupported reasoning can look as polished as grounded analysis
With RushDB
- RushDB keeps holdings, deals, memos, policies, and citations as connected records
- vLLM serves the chosen model inside the controlled environment
- Ontology and retrieval constrain prompts to available structure and evidence
- The backend can refuse unsupported questions instead of inventing answers
Graph intelligence on ingest
Incoming data becomes queryable graph context.
RushDB turns structured data into graph-ready context without a separate modeling pipeline. Structure already encoded in a nested payload is linked immediately. For flat records imported from scattered sources, relationship analysis can propose stable cross-source patterns.
01
Normalize as data arrives
Import JSON or CSV. RushDB infers property types and adds new fields to the live, queryable ontology without a schema migration.
02
Auto-link nested structure
Nested objects become connected records automatically, preserving the parent-child graph structure already encoded in your payload.
03
Enrich scattered sources
After flat imports or schema changes, analyze the project ontology. RushDB can suggest join patterns and semantic relationship types for your review.
Suggested relationship analysis requires an LLM configured for the project. Suggestions stay in draft form until you approve them, so inferred domain meaning never mutates the graph silently. You can also add explicit relationships through the SDK or API.
Review suggested relationship patternsData model
One flexible graph for the workflow.
Start with the payload shape your product already produces. RushDB stores it as Records, infers typed properties, and keeps nested or approved domain relationships queryable.
RushDB stores portfolio evidence and policy records; local inference receives only cited backend-selected context.
{
"portfolioId": "fo-2026",
"strategy": "growth",
"HOLDING": [{
"holdingId": "holding-ai-infra-co",
"allocationPct": 6.8,
"COMPANY": { "companyId": "northstar-compute", "sector": "ai_infrastructure" }
}],
"DEAL": [{
"dealId": "series-b-follow-on",
"stage": "diligence",
"INVESTMENT_MEMO": [{ "memoId": "memo-northstar-q2", "status": "approved" }]
}],
"POLICY": [{ "policyId": "ic-approval", "requiresHumanReview": true }]
}Working example
Ask a portfolio question. Return cited evidence or refuse.
The backend retrieves holdings and approved investment memos from RushDB, sends only cited evidence to the local vLLM endpoint, and requires record IDs in the response.
PORTFOLIO fo-2026
HOLDING ai-infra-co allocation: 6.8%
COMPANY "Northstar Compute"
INVESTMENT_MEMO "GPU supply concentration remains the primary risk..."
POLICY "Do not recommend trades without IC approval."
DEAL "Series B follow-on diligence"{
"labels": ["INVESTMENT_MEMO"],
"propertyName": "text",
"query": "AI infrastructure exposure, concentration risk, and follow-up diligence",
"where": {
"PORTFOLIO": { "portfolioId": "fo-2026" },
"status": "approved"
},
"limit": 8
}{
"answerability": "grounded",
"summary": "AI infrastructure exposure is concentrated in Northstar Compute.",
"citations": ["holding-ai-infra-co", "memo-northstar-q2", "policy-ic-approval"],
"unsupportedClaims": [],
"requiresHumanReview": true
}TypeScript SDK
Retrieve evidence first. Run local inference second.
RushDB supplies ontology, retrieval, relationships, and citations. vLLM runs the model inside the controlled environment. The thin UI only calls this backend workflow and is outside the case-study scope.
Implementation blueprint
Build the on-prem wealth-intelligence path.
Use this sequence to build a private backend for portfolio intelligence without turning the model into the source of truth.
- 01Deploy RushDB and vLLM inside the controlled environment
- 02Import PORTFOLIO, HOLDING, COMPANY, DEAL, INVESTMENT_MEMO, POLICY, and SOURCE records
- 03Create managed or external indexes for memo and policy text
- 04Retrieve ontology, holdings, policies, and cited memo chunks before inference
- 05Send only evidence payloads to vLLM and require citations or refusal
- 06Keep the thin UI out of the reasoning path; it only renders backend results
Build path
- Model portfolio intelligence as records and relationships, not prompt-only memory.
- Keep local inference behind an evidence-gated backend endpoint.
- Require answers to cite RushDB record IDs and source documents.
- Add human-review and investment-policy gates before any recommendation workflow.
Relevant docs
Read the exact primitives behind this pattern.
These links point to the RushDB docs pages that map directly to this case study: ingestion, labels, properties, values, SearchQuery, relationships, semantic search, MCP, or deployment.
Self-hosting RushDB
Deploy RushDB in a controlled environment and configure OpenAI-compatible embedding or LLM endpoints where needed.
Open docsSemantic search
Retrieve approved memos, policy text, and diligence notes by meaning before local inference.
Open docsAI overview and ontology
Use ontology so backend prompts only include fields, labels, and relationships that exist in the portfolio graph.
Open docsHow it works
Build the smallest useful workflow first.
01
Keep inference on-prem
Serve the model through vLLM inside the controlled network and send it only the evidence selected by the backend.
02
Build an evidence graph
Store holdings, companies, funds, deal memos, policies, diligence notes, and source documents as connected RushDB records.
03
Gate every answer
Load ontology, retrieve cited records, apply access filters, and require the model to cite evidence or refuse unsupported claims.
Know where it fits.
Grounded does not mean guaranteed correct
This pattern reduces unsupported answers by evidence-gating prompts and requiring citations. It does not guarantee financial correctness, investment suitability, or zero hallucinations.
Thin UI stays out of scope
The case focuses on the private backend: RushDB for evidence, vLLM for local inference, and policy gates for review. The UI should remain a thin client over those decisions.
Questions developers ask.
Next step
Start with one focused workflow.
Related use cases
Explore all use cases