# Thermo Fisher SampleManager LIMS AI Integration Without the Cloud > Source: https://netray.co/erp-help/private-ai-layer-thermo-samplemanager-lims-on-prem-architecture > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: ERP Help ## Short answer A private AI layer for SampleManager reads from a replica of the LIMS database, indexes SOPs, methods and attachments separately, and runs an open-weight model on local GPUs. Approved read tools sit behind an MCP or function-calling server, SampleManager roles are mirrored into every query, and any write goes through the supported application interfaces with human approval. ## Kind advanced ## System on-prem-apps ## Erp Thermo Fisher SampleManager LIMS ## Module Samples, Jobs, Results / Private AI layer ## Query SampleManager LIMS AI integration without cloud ## Alternate Queries - how to add AI to Thermo Fisher SampleManager on-prem - SampleManager LIMS local LLM - SampleManager LIMS natural language query ## Applies To Thermo Scientific SampleManager LIMS on Oracle, or on SQL Server where your release supports it. Confirm schema details in your own data model. ## Steps ### Roll out a private SampleManager assistant **Items:** - Agree the first ten questions with lab supervisors, for example overdue jobs, samples awaiting review, out-of-specification results by product, and instrument or analyst workload. - Provision a read-only copy of the LIMS database. On Oracle that may be a standby opened read-only, with licensing checked, or a scheduled export or restore into a reporting schema. - Create views that expose samples, tests, results, jobs, specifications and users in readable form, hiding configuration noise, and grant the AI service SELECT on those views only. - Index SOPs, methods, specifications and attached reports into a local vector and keyword index, keeping version, effective date and owning group on every chunk. - Serve an open-weight model with vLLM or Ollama on local GPUs and register the view-backed tools in an MCP server that the model can call. - Translate SampleManager roles and lab or product restrictions into row filters inside the tool server, and test with restricted accounts. - Turn on audit logging, start with read-only answers, then add drafts, and only later add approved actions through the supported interface. ## Sections ### How SampleManager exposes its data **Body:** SampleManager is a data-model-driven LIMS. Its database holds samples, the tests and results attached to them, jobs that group work, and large amounts of configuration: analyses, specifications, lists, users and roles. Business behavior is implemented in its own scripting layer, VGL (Visual Graphics Language), with tasks, rules and automation defined in the application rather than in the database. That matters for integration. Reading is straightforward from a replica, but writing around the application bypasses VGL rules, workflow state and the audit trail. For anything that changes data, use the documented application interfaces for your release. Ask Thermo Fisher or your implementation partner which web service or API layer your version includes. Attachments and reports are the second source. Certificates, chromatograms, instrument exports and scanned forms often sit in a file store referenced from the LIMS, and the assistant needs those paths to answer document questions. **Code:** -- Illustrative: SampleManager entity names are SAMPLE, TEST, RESULT; confirm columns in your model SELECT s.IDENTITY, s.STATUS, s.LOGIN_DATE FROM SAMPLE s WHERE s.STATUS = 'P' -- example status code, check your config AND s.LOGIN_DATE < SYSDATE - :days ORDER BY s.LOGIN_DATE; # Tool definition exposed to the model (MCP style) { "name": "overdue_samples", "description": "Samples in progress older than N days, filtered to the caller's labs", "inputSchema": { "type": "object", "properties": { "days": { "type": "integer" } } } } ### Reference architecture **Body:** Replica, curated views, tool server, retrieval index, local model, audit store. The tool server is the only component that talks to the database, and it holds the caller's identity. The model is told which tools exist and never receives database credentials. Change data capture keeps the document index fresh. A nightly job is enough for SOPs. For result-related text such as comments and deviation narratives, an hourly sync is typical. Embeddings and reranking also run locally. For an air-gapped site, load weights and containers through a transfer station, mirror package repositories internally, and disable all telemetry. Place the AI hosts in their own network segment with firewall rules that allow only database replica and file share reads. ### Permissions and data segregation **Body:** SampleManager enforces roles and, at many sites, restricts what a user can see by lab, department or product. The assistant must not widen that. Resolve the user's SampleManager identity at sign-in, look up their role and restrictions, and apply them as filters in every tool query. Do the same for documents: tag each chunk with the group that may read it and filter retrieval before the model sees text. Never rely on prompt instructions to hide data. Add a canary test suite that runs nightly: restricted users asking for data outside their scope must get nothing, and failures page the owner. ### Write-back and validated environments **Body:** Keep write-back narrow and approved. Good candidates are drafting a login request, a deviation text or a comment, which a person reviews and submits in SampleManager itself. Poor candidates are result entry, review and authorization. Most SampleManager customers run in regulated settings. Document intended use, run a risk assessment, and keep the assistant advisory. A read-only, fully logged tool with human review fits more easily into your computer system validation approach than autonomous actions. ### Sizing, phases and limits **Body:** For 100 to 400 named users and a few questions each per day, one or two data-center GPUs usually carry a 14B to 70B class model, with headroom for peak mornings. Index size depends on document volume; a few hundred thousand pages is tens of gigabytes at most. Run the calculator with your numbers. Plan eight to sixteen weeks: weeks one to four for replica, views and tools; then retrieval over documents; then evaluation with real questions; then drafts. Do not skip the evaluation set. Limits: heavily customized data models need extra mapping, language models are poor at multi-step numeric analysis, and if supervisors only need a fixed report, SampleManager's own reporting is the simpler answer. ## Pitfalls - Writing straight into LIMS tables, which skips VGL rules, workflow state and audit trail. - Exposing every table to the model instead of curated, documented views. - Ignoring lab and product restrictions because the pilot users are all administrators. - Indexing documents without effective dates, so superseded SOPs are quoted as current. - Skipping a real-question evaluation set and judging the assistant by demos. ## Ai Angle For SampleManager sites, Netray's custom build delivers the replica, views, tool server, document retrieval and local model, tuned to your configured data model. ERPray covers the ERP side of the same questions, for example linking a failed lot to supplier receipts, while DataRay handles file-share documents on-prem. ## Faq ### Can we query SampleManager in plain English? **Answer:** Yes, through curated tools rather than free SQL. The model picks a tool such as overdue samples, fills parameters from your question, and the tool server runs a tested query with your permissions. Free-form SQL generation is possible later but needs guardrails. ### Does this need SampleManager upgrades? **Answer:** Not for read-only access through a replica. Only write-back through application interfaces depends on what your version provides. Confirm the interface availability with Thermo Fisher or your implementation partner. ### How are VGL customizations handled? **Answer:** VGL rules stay where they are. Reads ignore them, which is why the views must expose the fields as the lab understands them. For writes, calling the application interface keeps VGL rules and audit trail in force. ### Is cloud AI ever acceptable here? **Answer:** Sometimes, for non-sensitive text such as public method references. If sample data or proprietary formulations are involved, or contracts forbid it, keep the model local. The architecture works for either, but local is the default for this data. ## Search terms samplemanager lims ai, thermo fisher samplemanager integration, samplemanager lims local llm, samplemanager natural language query, lims ai on premise, samplemanager vgl integration, private ai lims oracle, samplemanager sop search ai --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json