Knowledge Query Platform for Health Plan Alliance.
Turning processed knowledge sources into grounded, citation-backed answers. Health Plan Alliance needed more than a public chatbot — Halley AI deployed a controlled, single-tenant knowledge query platform that helps teams search approved content, review supporting evidence, and refresh the processed corpus through a managed workflow.
From scattered knowledge to governed search.
5-minute read
Health Plan Alliance works with knowledge that lives across documents, tables, summaries, operational references, and program-specific materials. The value is high, but the retrieval problem is real: people need accurate answers without manually opening file after file or trusting an opaque answer with no source trail.
Halley AI built a browser-based knowledge query platform around a processed corpus. Source material can begin in many file formats, then moves through a normalization pipeline that extracts text, tables, summaries, and metadata into a structured bundle. The platform ingests that bundle, chunks it, creates embeddings, and stores the searchable vectors in a local PostgreSQL database with pgvector.
The Challenge
High-value knowledge was hard to query reliably
Health plan teams often need answers that depend on the exact wording, source, and context of internal knowledge. A generic assistant can sound confident while missing the evidence trail that users need for trust and review. A static document library can preserve the record but still force staff to search manually, compare files, and reconstruct context by hand.
The platform needed to support practical discovery while respecting operational boundaries: browser access, role-aware administration, controlled refreshes, and clear separation between source material, vector storage, and answer synthesis.
The Solution
A controlled retrieval-augmented knowledge platform
Halley AI deployed a Knowledge Query Platform that starts with processed corpus bundles rather than raw browser uploads. The pipeline organizes extracted text, tables, executive summaries, flags, questions, and manifests into a stable structure that the application can ingest and audit.
- Processed-corpus ingestion: The application ingests curated bundles with extracted text, tables, summaries, and metadata instead of asking users to upload raw documents into a public chat interface.
- Hybrid retrieval: Lexical search and vector search work together so users can find both exact terminology and semantically related evidence.
- Grounded answer synthesis: The answer layer uses retrieved local evidence and returns source-linked responses. Answer calls are configured with
store=false. - Analyst Mode: Users can inspect evidence transparency without exposing internal reasoning output, helping reviewers understand which chunks supported the response.
- Admin refresh: Authorized administrators can review ingestion status and trigger refreshes from configured processed-corpus bundle paths.
Architecture
Designed for healthcare-adjacent data boundaries
The implementation intentionally avoids OpenAI File Search, OpenAI vector stores, and full-corpus upload to OpenAI-managed storage. Embeddings are generated for retrieval, but returned vectors are stored in the client platform's PostgreSQL/pgvector database. Citations are validated against locally retrieved chunk identifiers before being shown to the user.
Production architecture is designed around a single-tenant browser application with xCatalyst OIDC authentication, Viewer and Admin roles, and no API keys or secrets exposed in the browser. That makes the platform a fit for custom implementation conversations where controlled access, evidence review, and deployment boundaries matter.
Compliance note: This case study describes a controlled custom implementation pattern. HIPAA scope depends on the deployed environment, data-handling rules, vendor contracts, BAA coverage, access controls, retention policy, monitoring, and operating procedures. It is not a blanket statement that every public Halley AI assistant or marketing widget is HIPAA compliant.
Operational Impact
Answers with sources, not just chat
The platform gives users a practical way to query complex knowledge while keeping the evidence visible. Instead of treating AI as a standalone chatbot, the system behaves like a controlled retrieval layer over a curated knowledge base.
- Faster answer discovery: Users can ask natural-language questions across processed knowledge sources instead of manually searching each file.
- Higher trust: Source-linked responses and Analyst Mode make it easier to review why an answer was produced.
- Repeatable refreshes: Administrators can update the processed corpus through a governed path without changing the user experience.
- Production readiness: UAT checks cover authentication, role boundaries, citation validation, low-evidence behavior, performance smoke tests, and safety controls.
Conclusion
A practical pattern for controlled AI knowledge access
Health Plan Alliance's Knowledge Query Platform shows how Halley AI can move beyond embedded website assistants into client-specific AI systems. The same core idea applies to associations, healthcare operations, training libraries, member support teams, and internal knowledge centers: normalize the content, store searchable vectors under client control, retrieve evidence, and generate answers with citations.
For organizations evaluating AI in sensitive workflows, the important question is not whether a chatbot can answer a sample question. The question is whether the system can respect data boundaries, surface evidence, refresh knowledge predictably, and fit the way the organization actually operates.
Need a controlled knowledge query platform?
A guided implementation that respects your data boundaries — with evidence, citations, and managed refresh.