Conventional public search
Keyword and metadata retrieval from approved public pages. No model, account, cookie or external AI provider is required.
Responsible AI
Responsible AI starts with evidence, privacy and a clear right to abstain.
NovaPharm uses the smallest justified technology for each task. Public search is retrieval-led, private by default and unable to access secure records. Internal AI is provider-disabled by default and remains subordinate to human accountability.
Operating principle
Retrieve approved evidence before considering generation.
The public experience does not use a general-purpose chatbot. It finds controlled NovaPharm pages and Insights, applies deterministic safety rules, shows exact supporting passages and abstains when the approved corpus cannot support an answer.
Semantic similarity helps find a passage; it does not establish that the passage is true beyond its source, current beyond its review date or sufficient for a regulated decision.
AI maturity
Every capability has a visible status and a prohibited-use boundary.
Status is based on implemented and tested behaviour, not a business-plan ambition.
Private on-device semantic retrieval
A small NovaPharm lexical-semantic retrieval model is downloaded only after explicit activation. Queries remain in the visitor's browser.
Governed review assistance
Bounded prototypes can flag possible claims contradictions, document gaps and enquiry categories for human review. They cannot approve records or regulated decisions.
Forecasting and traceability research
Demand planning and stronger batch traceability require governed data, validation, monitoring and accountable owners. No performance or blockchain-deployment claim is made.
Medical and autonomous regulated decisions
Public AI must not provide diagnosis, treatment, dosage, medicine selection, patient advice, availability claims or autonomous regulatory, quality, safety, credit or release decisions.
Citation architecture
An answer is a view over sources, not a new authority.
Permitted questions are searched across deterministic chunks derived from approved canonical pages. The interface returns source title, URL, heading, supporting passage, date, capability boundary and evidence status.
- 01Policy first
Refuse medical, safety, private, financial, live-stock and prompt-injection requests before retrieval.
- 02Approved corpus
Search only registered public pages, published Insights and approved leadership profiles.
- 03Exact passage
Build a concise extractive answer and retain the underlying passage and canonical source.
- 04Abstain
Say “I could not verify that from NovaPharm's approved public information” when support is insufficient.
Privacy by default
The public evidence assistant does not need to know who you are.
Conventional and optional semantic retrieval are designed without query logging, external inference, microphone access or transcript retention.
- Query text
- Processed in the browser and not sent to an external AI provider or stored by the assistant.
- Model activation
- Optional, explicit and cancellable. The size is shown before download.
- Browser storage
- Optional semantic assets may be cached locally; the interface provides a clear-cache control and falls back when storage is denied.
- Analytics
- No query or answer text is measured. Only non-sensitive aggregate interaction events may be considered after consent and approval.
- Secure records
- Portal records, customer data, uploaded documents, credentials and private business-plan content are not in the public corpus.
- Medical information
- The assistant refuses patient, treatment, dosage, adverse-event and product-defect content and directs users to controlled routes.
Model governance
A small local retrieval model, not a public generative model.
NovaPharm Evidence Vector v1 creates deterministic sparse embeddings from approved English vocabulary. It has no neural weights, no remote model dependency and no ability to generate prose independently of retrieved passages.
The model, source manifest, corpus hash, asset checksums, licence decision and limitations are recorded in the repository. The optional model is lazy-loaded in a Web Worker so the normal website experience remains independent of AI.
Use the no-JavaScript search directoryInternal gateway
Private AI remains off unless an authorised use case and provider are configured.
The secured Node architecture defaults to provider `none`. A loopback-only Ollama adapter is available for local development, with server-side role checks, source restrictions, redaction and audit metadata that excludes prompt bodies.
No internal AI route can write a production record, approve a supplier, release a batch, make a regulatory decision or cross customer boundaries. Any future provider requires security, privacy, legal, model and use-case approval.
Governance references
Controls are aligned to current primary guidance.
These sources inform the governance approach; they do not certify NovaPharm or replace a task-specific legal and regulatory review.
Report an issue
Challenge an answer, source or boundary.
Send the public page or source URL and a non-sensitive description. Do not submit patient data, private documents, passwords or confidential commercial information.