Read summarized version with
About project
Working time:
Since 2026
Industry:
Life sciences
The service:
Agentic AI
Overview
The customer is a Canadian life sciences company that develops research tools and reagents for laboratories. Its scientists and application specialists support customers worldwide and rely on years of protocols, product specifications, application notes and internal study reports.
The Challenge
The knowledge existed, but finding it took hours. Content was spread across file shares, a document management system and a product database, and search was keyword based.
1. Scattered Knowledge
Protocols, specifications and reports lived in separate repositories, each with its own search.
2. Repeated Work
Scientists repeated experiments that had already been run because past reports were hard to find.
3. Slow Customer Answers
Application specialists answered the same technical product questions from scratch.
4. Untraceable AI Answers
A general chat tool answered without citations, so scientists could not check it and did not trust it.
Technology & Approach
We built search agents that answer only from the company’s governed scientific content and show where every statement comes from.
Implementation
Phase 1: Corpus and Questions
Inventoried sources, collected 300 real questions from scientists and application specialists, and agreed access rules by team and document class.
Phase 2: Ingestion and Retrieval
Tuned ingestion and chunking for protocols, tables and figures with Amazon Textract, indexed in Amazon OpenSearch Serverless through Amazon Bedrock Knowledge Bases.
Phase 3: Agents on AgentCore
Agents on Amazon Bedrock AgentCore answer with citations and query the product database through governed tools exposed by AgentCore Gateway.
Phase 4: Citation Checks and Pilot
Every answer is checked against its cited passages and unsupported statements are flagged, piloted with the application support team before extending to R&D.
Customization & Client Adaptation
Scientific content needs different handling from ordinary business documents, so we adapted each layer to it.
Scientific Document Handling
Chunking keeps protocol steps, reagent tables and figure captions together, so answers do not lose their context.
Access by Team
AWS IAM Identity Center groups decide which document classes each team can query, and the same rules apply inside the agent.
Answers You Can Check
Each answer lists its sources with direct links, and statements without support are shown as such instead of as fact.
Structured Data Through Tools
Product specifications come from the product database through a governed tool, not from copies of documents.
If you find this case interesting, we recommend taking a closer look at
The Results
Application specialists now answer common technical questions in minutes rather than hours, and nine in ten pilot answers were rated correct and fully cited by the scientists who reviewed them.
Long-Term Value
- One governed corpus that new sources join without a new tool.
- A question set that measures accuracy as the corpus grows.
- Agents that can take on new tools, such as lab data systems, under the same access rules.
Business Impact
- Past experiments and specifications are found instead of repeated.
- Customers get faster, consistent technical answers.
- Scientists trust answers because they can check every source.
