ConsensBot

Document lookup in a regulated plant

In a GMP environment, where the evidence sits matters before whether the answer is right. Staff look up SOPs, quality records and electronic documents in separate systems and compare them by eye. It takes time, and different people sometimes come back with different evidence.

Attaching a general-purpose LLM makes matters worse, because it will invent plausible content that is not in the documents. In regulatory work, a plausible wrong answer is worse than no answer.

Answer verification

ConsensBot adds a verification step after generation.

The main claims and cited evidence in each answer are automatically checked against the source text. If the evidence is thin or conflicts with the original, the system searches and generates again; if it still cannot be confirmed, it replies “not confirmed”. This check runs on every answer rather than a sample, and takes 0.7 seconds on average.

Processing sequence

1. Intake — The question’s intent and key terms are analyzed and classified by type (document lookup, procedure guidance, record check), and the search scope is set across EDMS, QMS, the internal wiki and business databases.
2. Gathering evidence — Relevant material is collected from SOPs, MBRs, SPECs, quality records and change or deviation records. Keyword and semantic search run together across body text, tables, metadata and structured records, narrowing to the latest version, document number, table field and record.
3. Answering — The answer is written from the selected evidence, with the document name, number, version and relevant fields attached. It only reaches the user after passing the full verification described above.

Search scope

Internal SOPs and regulations, quality records and electronic documents, through one interface. Hybrid retrieval combining dense and sparse embeddings means terms do not have to match exactly.

It is not only lookup. Comparing several documents side by side, pinpointing what changed between revisions, and aggregating quality records by condition are all handled in the same place.

It connects to QMS, EDMS and other document management systems.

On-premise operation

Every component — vLLM, the vector database and the embedding models — runs on customer servers. No external API is called, so regulatory documents and quality records never leave the company.

Deployment is either on-premise or a dedicated customer cloud (VPC). IMPIX holds ISO/IEC 27001 and GS Certification Grade 1.

Knowledge accumulation

When someone on the floor reports that an answer should be corrected, that feedback is reviewed by an administrator and accumulated into a knowledge wiki. The judgement the organization has built up survives staff turnover.

In production

Quality and production staff in pharmaceutical and bio manufacturing query the system conversationally; it searches, compares and verifies the relevant documents and answers with citations.

The deployment searches 1,527 regulatory documents, 146,919 quality records and over 63,000 electronic documents, and every generated answer passes automatic evidence verification. 119 regression tests run continuously to confirm that quality is not degrading.

Adoption requirements

Deployment and servers — Built on-premise or in a private cloud according to your security policy. It needs an internal network path to the business systems and document stores, a GPU server for the LLM and embedding models, and server capacity for the API, search and vector database. Where external network access is restricted, the procedure for bringing in and updating models and packages has to be defined in advance.

Document readiness — SOPs, MBRs, SPECs and other standards, QMS records and the internal wiki form the evidence base. Metadata such as document number, type, version, effective date and authoring department must stay linked to the source text, and duplicates, withdrawn documents and superseded versions must be distinguishable. Tables and body text need to be extracted into a searchable form.

Integration — Connects to EDMS, QMS, the internal wiki and electronic document stores, plus any required business databases. API or database queries are the default, with file and batch transfer where a system calls for it. Changed data is re-collected and re-indexed periodically, and identifiers such as document number and version are preserved from the source system.

Permissions and operation — Roles are split between general users, administrators and domain reviewers. Users read answers and evidence and submit feedback; administrators and domain owners review corrections and negative feedback, then approve or reject them. Only approved content reaches the knowledge wiki and ontology, and the change and approval history and the evidence behind each answer must remain traceable.

Contact us and we will provide a free PoC environment.