Ingestion pipelines per source
Connectors for file stores, wikis, ticketing systems or databases that extract text, tables and metadata, handle scanned PDFs with OCR where needed, and record owner, version and access list for every item.
Service · AI
Commission a retrieval-augmented generation system that answers from your own manuals, policies, tickets or contracts, shows where each answer came from and never surfaces a document the asker is not entitled to read. We build ingestion, indexing, retrieval and the answer layer, then prove them on questions your users really ask.
Access rules enforced at retrieval time
Every answer linked to its source passages
Retrieval and answers graded separately
Deliverables
Connectors for file stores, wikis, ticketing systems or databases that extract text, tables and metadata, handle scanned PDFs with OCR where needed, and record owner, version and access list for every item.
Documents split along their natural structure, such as headings, clauses or ticket turns, enriched with titles and context, and stored in a vector index alongside a keyword index for exact terms like part numbers.
The signed-in user’s identity and group memberships travel with every query, filtering candidates before ranking so restricted content never reaches the model, the cache or the citation links.
Generation that quotes or links the passages it relied on, declines when the retrieved material does not support an answer, and flags conflicting versions instead of silently picking one.
Real questions collected from users or support logs, each tagged with the documents that should be found and the answer a subject expert accepts, used to score retrieval and generation independently.
Early decisions
A Postgres extension such as pgvector keeps everything in one database you already back up; a dedicated vector service scales differently. Corpus size, query volume and your operations team decide which.
Pure embeddings miss exact codes and names; pure keyword search misses paraphrase. We usually combine both and test whether a reranking model earns its added latency on your questions.
Too small and context is lost; too large and irrelevant text dilutes the answer. We compare boundary strategies on the graded set rather than picking a token count from a blog post.
Copying permissions into the index is fast but can go stale; checking the source system at query time is accurate but slower. We choose per source, based on how often its permissions change.
Edits, removed documents and departed employees must disappear from every serving copy. We set re-sync intervals, deletion propagation and how long cached answers may live.
Language coverage, domain vocabulary, licence and the cost of re-embedding the whole corpus later all matter. We benchmark candidates on your text before committing.
Engagement
We list every repository in scope, who owns it, how access is granted and how deletions happen, and collect a first batch of real questions with expected sources.
Ingestion and search are built first and scored on whether the right passages appear near the top; only when retrieval is solid do we tune the answering step.
Citations, refusal behaviour, feedback buttons and the interface or API that delivers answers are added, then tested with a pilot group who log bad answers.
Scheduled re-indexing, deletion handling and quality dashboards run in production while we tune; your team then receives pipelines, index configuration and the graded question set.
Fit
Staff or customers lose time searching wikis, shared drives and old tickets for answers that exist somewhere but are hard to find.
In support, compliance, legal or engineering contexts, a reply is only useful if the reader can check the paragraph it came from.
Policies, prices or product specs are updated often enough that fine-tuning a model on them would leave it out of date within weeks.
Alternatives
If the hard part is dialogue design, channel coverage and handover to staff rather than finding documents, our conversational AI work leads.
Open registerWhen a request should end with a record updated or a refund issued, an agent with scoped tools fits better than retrieval alone.
Open registerBehaviour that cannot be supplied as reference text, such as a specialist output format, may need tuning at the model layer instead.
Open registerA language model cannot give a good answer from the wrong passages. When users complain that a RAG assistant is “making things up”, the usual cause is that retrieval returned weak or irrelevant text and the model filled the gap. That is why we score retrieval on its own, using recall at a fixed depth, before spending any effort on prompts.
Improving retrieval often comes down to unglamorous work: better text extraction from PDFs, chunks that follow document structure, metadata such as product line or region, and a keyword index for terms embeddings handle poorly.
An internal assistant that reveals salary bands or board minutes to the wrong employee is worse than no assistant. Access rules must be applied before ranking, carried into caches and reflected in citation links. We also test revocation: when someone leaves a group, their next query must no longer reach that group’s files.
A fluent reply can still be unsupported. We grade three properties separately: whether the right sources were retrieved, whether the answer is faithful to them, and whether it actually addresses the question. Model-based graders speed this up, but we calibrate them against expert judgements so their scores mean something.
Documents are edited, superseded and deleted constantly. An index refreshed once at launch will slowly fill with outdated policy. Sync jobs, deletion handling and a visible “last updated” date on cited sources keep users’ trust, and dashboards show when a source has stopped syncing.
Retrieval suits questions whose answers exist somewhere in your content. It does not teach a model a new skill, and it does not take actions. When a project needs either, we combine retrieval with model tuning or with agent tooling, and our sibling services handle those parts.
FAQ
The number and messiness of sources is the largest factor: scanned documents, inconsistent formats and complex permission models all add work. Corpus size affects embedding and storage costs, and a strict accuracy target means a larger graded question set and more tuning.
Access to source systems and their permission data usually sets the pace, followed by how fast subject experts can grade sample answers. Teams that can provide connectors and reviewers early move through the retrieval phase much faster.
Ingestion and sync code, index schemas and configuration, the retrieval and answer services, the user interface or API, the graded question set with scoring scripts, and documentation on re-indexing and adding a new source.
Yes. The software, configuration and evaluation material we create are assigned to you on payment. Your documents never become ours, and the index lives in infrastructure you control or nominate.
We agree separate thresholds for retrieval, meaning the right source appears in the top results, and for answers, meaning experts rate them as correct and supported. A permission test proving restricted documents stay hidden is a mandatory part of acceptance.
Rewriting or curating your content, fixing permission errors in source systems, and translating documents are outside scope. We report contradictions and gaps we find so content owners can fix them at the source.
Related services
When source systems need cleaning, deduplication or new export feeds before they can be indexed reliably.
Open registerFor tuning a pgvector deployment, partitioning a large corpus or adding indexes alongside existing tables.
Open registerFor tracking answer quality over time and gating changes to embeddings, prompts or rerankers.
Open registerPrimary sources
Dated official documentation, standards, and research that support the factual claims on this page.
Open-source vector similarity search for PostgreSQL.
A write-up of hybrid search and chunk context techniques for retrieval.
How to read this page
Scope, cost and timeline figures are planning guidance until confirmed in a written proposal. Diagrams and screens are illustrations, not client deliverables, and source-code rights and handover follow the signed agreement. See the delivery process and the full disclaimer.
Next decision
Define outcomes, constraints, evidence, rights and handover before delivery begins.
Commercial rights, repositories, environments, documentation, acceptance and handover remain contract-defined.