Your assistant gives poor answers
It retrieves the wrong passage, cites an old version or answers without a source.
Mintera Knowledge · Data / AI consulting & engineering
Scattered data.
Usable knowledge.
We assess your sources, prepare your documents and build retrieval for your AI tools. You receive documented architecture, governance rules and a knowledge base evaluated against your team’s questions.
Your data remains fragmented
The same information lives in document systems, tickets, shared files and business tools.
The pilot does not scale
Updates, permissions, costs and maintenance have not been addressed.
Three engagement scopes
Defined work.
Deliverables you retain.
Each stage can be commissioned separately. The proposal specifies sources, connectors, volumes, responsibilities and acceptance criteria.
Data assessment & AI strategy
Decide where to invest and what to start with.
What we do
- Workshops with business, IT and data owners.
- Inventory of sources, formats, volumes, permissions and current flows.
- Sampling for duplicates, scans, versions, sensitive content and stale data.
What you receive
- Source and data-flow map.
- Use-case matrix: value, feasibility, risk and cost.
- Target architecture, prioritised backlog and staged budget.
Decision: an agreed scope, business owner and success criteria.
Data preparation & knowledge base
Build the pipeline from source content to useful context.
What we do
- Scoped ingestion, text extraction and OCR where required.
- Metadata, deduplication, chunking and traceability to the source document.
- Indexing, retrieval and permission propagation; updates and deletions.
What you receive
- Metadata dictionary and quality rules.
- Documented pipelines and index schemas.
- Business questions with expected source evidence.
Decision: the right users can retrieve the right sources.
Evaluation & production readiness
Check reliability and organise operations.
What we do
- Compare retrieval strategies against the same test set.
- Test answers, citations, permissions and missing-information cases.
- Measure cost, latency and monitoring needs.
What you receive
- Evaluation report and gap register.
- Deployment, monitoring and reindexing plan.
- Operating documentation and team handover.
Decision: agreed thresholds are met, or limitations are explicit.
Recurring operations, licences, model calls and business-agent development are costed separately. A pilot scope tests design choices; it does not replace governance across the full data estate.
How we make data usable
AI receives useful context.
Not your entire file server.
A terabyte of files does not map directly to tokens. Retrieval selects relevant evidence for a question; access rules are applied before passing it to the model.
- 01
Ingest
Document systems, files, tickets and business exports: scope and permissions checked.
- 02
Prepare
Text, tables, versions, metadata and provenance retained.
- 03
Retrieve
Keywords + semantic search, filters and passage ranking.
- 04
Answer & check
Authorised context, inspectable references and abstention when needed.
Retrieval-augmented generation (RAG) is an option, not a requirement. For structured numerical data, a controlled query or calculation can be preferable to retrieving text passages.
The engineering details that matter
What we investigate.
What we test.
01Extraction & quality
We test native PDFs, scans and tables on a representative sample. Unreadable documents are quarantined with an explicit reason rather than silently ignored.
Check: processing coverage, extraction errors and corpus coverage.
02Metadata & versions
Each document has an identifier, source, owner, validity date, version and classification. Chunking preserves headings, page references and links to the original.
Check: retrieve the applicable version and its provenance.
03Retrieval & ranking
We compare keyword, vector and hybrid search against your questions. Reranking is added if it improves quality at an acceptable cost and latency.
Check: expected evidence appears in the first results (Recall@k).
04Access & confidentiality
We model users, groups and document permissions. Filters apply during retrieval; a prompt instruction is not an access control.
Check: cross-role tests, permission revocation and forbidden sources.
05Freshness & deletion
The pipeline detects new, changed and deleted documents. Failed processing can be resumed and content reindexed after schema or embedding-model changes.
Check: freshness delay, deletion propagation and no stale ghost versions.
06Evaluation & operations
We build questions with expected answers and evidence, validated by business users. Tests also cover conflicting documents, out-of-scope questions and malicious instructions inside source content.
Check: supported citations, accepted answers, appropriate refusals, p95 latency and cost per query.
Three practical applications
A shared foundation.
Different business criteria.
SaaS vendor · Support
- Sources
- Resolved tickets, product documentation and release notes.
- Structuring work
- Link evidence to the product version; exclude another customer’s confidential content.
- Benefits to measure
- Search time, replies approved without rework and avoidable escalations.
Consulting firm · Proposals
- Sources
- References, engagements, methods and previous proposals.
- Structuring work
- Separate reusable templates from commitments and figures specific to a former client.
- Benefits to measure
- Preparation time, references found and corrections before submission.
Industry / data centre · Operations
- Sources
- Procedures, manuals, incidents and intervention reports.
- Structuring work
- Link equipment to the applicable procedure and flag conflicting versions.
- Benefits to measure
- Time to find a valid instruction and incidents involving outdated versions.
Measurement example, not a customer result
Across 400 requests a month, reducing search time from 12 to 7 minutes frees about 33 hours. This is a hypothesis to test; verification time, maintenance and service costs must be deducted to calculate net benefit.
Useful answers
Before you start.
Which sources can you assess?
Document systems, shared drives, SharePoint, CRM/ERP exports, ticket systems, technical documents and object storage. APIs, export formats and permissions are checked before committing to a connector.
Must all data move to Mintera?
No. Some sources remain in your tools. Scoping specifies what is read remotely, copied, indexed or stored. Storage, index and model locations, providers and data flows are agreed before implementation.
How is the proposal priced?
Assessment, preparation, integration and operations are priced separately. Source count, formats, permissions, quality and update frequency drive effort. A representative sample helps estimate work before large-scale processing.
What does my company retain?
The contracted deliverables: maps, schemas, quality and access rules, pipeline documentation, evaluation set and operating procedures. Rights to use custom developments and third-party dependencies are specified in the proposal.
Your Mintera Knowledge lead
Benjamin Jornet
Data & AI value creation · 8 years at Sopra Steria
Benjamin brings eight years of experience at Sopra Steria and expertise in data challenges and creating value with AI. At Mintera, he is your lead for connecting business requirements with data preparation and integration choices.
His role on your project: challenge the scope, assess sources, define quality criteria and coordinate technical decisions with your team.
Discuss my project with BenjaminStart with your sources.
And a task to improve.
Tell us about your tools, relevant data and desired outcome. We will define the work needed and a scope that can be priced.