Documents, contracts, manuals, code and internal communications accessible via chat with source citation. No hallucination.
Company knowledge sits across SharePoint, Drive, Confluence, PDF manuals, old Slack threads and the head of whoever has been there for fifteen years. Getting hold of it in time is one of the quietest drains on productivity: the engineer who works out an issue that was already solved two years ago, the account manager who cannot find the terms agreed with a client, the quality lead who finds out too late that someone worked from a superseded revision of a procedure. The search box in the document manager does not fix it, because it only returns what literally contains the word you typed, and real company documentation is written by different people, in different years, in different words.
A well-built RAG system turns that document base into a conversation: embeddings that capture the meaning of each chunk, a vector store that inherits permissions from the source system, a re-ranker that reorders candidates before anything is written, and a language model that composes the answer. What separates a useful RAG from a dangerous one is traceability: every answer cites the document, page and section it came from, and the person asking can open it and check. Where someone signs off a report or a quote on the strength of that information, an answer with no source is worth nothing, so we treat citation as a design requirement rather than something bolted on at the end.
Across the RAG deployments managed by Summum IA, answer accuracy reaches 91% with 100% citation traceability, on a corpus of 14,000+ technical documents. New engineer onboarding drops from 12 weeks to 4. Methodology and details available on request.
This agreement establishes a validity period of 24 months, automatically renewable unless terminated by either party with a minimum notice of 60 calendar days.
The contract renews automatically every 24 months, unless notice is given 60 days in advance.
Source · contrato-proveedor.pdf · page 4 · §3Before anything is indexed we map the corpus as it actually is: which repositories exist, who can see what, and what state the documentation is in. We flag superseded documents, duplicates where two versions are both live, and scanned PDFs that need OCR before they can be read at all. The output is a list of sources that go into phase one, the ones that stay out, and the clean-up work to do first, which shapes the final result more than anything else.
Each document is chunked with a strategy suited to its type: a contract is not split the way a maintenance manual or an email thread is. Chunks become embedding vectors stored in the vector store, Qdrant or Pinecone depending on the client's infrastructure. Permissions travel with the chunk: if someone cannot open a document in SharePoint, they cannot pull its contents through the chat either. Incremental ingestion is set up on that same foundation.
A query resolves in two stages. Candidate chunks are retrieved first, combining semantic search with exact matching, which you still need for product codes, case numbers and regulatory references. The re-ranker then scores them against the actual question and only the best ones reach the model, which writes the answer under an obligation to cite document, page and section. When the retrieved context does not support an answer, the system says so instead of filling the gap.
A RAG system is not delivered and forgotten. We leave a dashboard showing real queries, which ones resolve with a citation and which ones come back with nothing useful, since those point straight at the gaps in the corpus. That list gets reviewed with the internal owner, missing documents are fixed or added, and the index is refreshed. Continuous ingestion picks up new material without opening another project. Without that review, the system decays as fast as the documentation ages.
The operational detail: what we deliver as part of the engagement and what we keep active afterwards.
Vector store with auth
Vector index on Qdrant or Pinecone, chosen for the client's infrastructure and data residency requirements, with source-system permissions replicated chunk by chunk: each user retrieves only what they could already open in SharePoint, Drive or Confluence.
Fine-tuned chunking
A different splitting strategy per document family: by clause for contracts, by section for manuals and procedures, by message for email threads, with overlap between chunks so that neither an idea nor a table is ever cut in half.
Re-ranker for precision
A second pass over what was retrieved: a reordering model scores each chunk against the specific question and drops the ones that only matched on general similarity, so the language model writes from the passages that are relevant rather than whatever the index returned first.
Mandatory source citation
Every answer arrives with its source document, page and section, linked so they open in the repository. When the retrieved context does not support an answer, the system says so instead of filling in from the model's general knowledge.
Continuous ingestion
New documents and revisions of existing ones are indexed incrementally, on an agreed schedule and without rebuilding the whole index. Withdrawn versions are removed from the index so that nobody is handed a procedure that has been superseded.
Feedback loop
A dashboard of real queries, with useful or not useful marking by the people using it. Questions that come back unanswered expose the gaps in the corpus and turn into specific work: documentation to write, chunking to adjust, ranking to tune.
The project runs remotely from start to finish: the corpus audit, the indexing and the move into production all happen against the client's own repositories, so where the company is based does not constrain the work. Summum IA is part of Grupo Summum, with offices in Castilla y León (Valladolid, Burgos, Aranda de Duero and Palencia) and in the Canary Islands, and works from there on projects across Spain.
Gipuzkoa and San Sebastián are a case of their own within that coverage: precision engineering and machine-tool firms in the province produce technical manuals, product certificates and commercial correspondence in several languages, exactly the kind of corpus where keyword search falls short. For that profile we keep a dedicated page covering the local industrial base: AI consulting for companies in Gipuzkoa and San Sebastián.
We have no office open in the Basque Country and do not advertise one: the model is remote, with occasional visits agreed inside the project itself and kept for team training or go-live.
GDPR if the corpus contains personal data. AI Act if high risk.
RAG is AI, governance and access rights.
RAG for specific knowledge. Fine-tuning only if style or jargon justifies it.
Corporate auth respected.
Up to 100k+ in real projects.
Almost always because of the corpus, not the model. Four failures come up again and again: two live versions of the same procedure, so the system cites the wrong one; scanned PDFs with no text layer, which the index never reads; uniform chunking that cuts a table or a clause in half; and questions that span several documents, where retrieval brings back the closest match rather than the right one. None of these is fixed by switching language model. They are fixed by auditing the corpus before indexing, tuning chunking per document type, adding a re-ranking stage, and measuring with real questions from the business instead of lab examples.
It depends on the architecture, and that is decided at the start of the project. The vector index and the documents stay on the infrastructure the client agrees to. What travels to the language model are the chunks retrieved to answer that specific question, never the full repository. Where the sensitivity of the material demands it (personal data, drawings, case files, anything under a confidentiality agreement) we assess options with EU data residency or with the model deployed on the client's own infrastructure, at the cost of higher operating overhead. If the corpus holds personal data, the processing is documented under Regulation (EU) 2016/679.
Three concrete things. Someone who knows the documentation and can decide which version of each document is the correct one: that is usually the real bottleneck, more than the technical work. Access to the source repositories and to the corporate identity system, so permissions are inherited instead of reinvented. And a list of thirty to fifty real questions that currently take time to answer, which serves as the test bench and decides whether the system is ready for production. Architecture, connectors, indexing and the operations dashboard come from Summum IA.
Not if the goal is for the model to know what your documentation says. Fine-tuning adjusts how a model answers, but it melts the knowledge into the weights: you cannot cite a source or tell where a statement came from, and updating one procedure means running the training again. With RAG you reindex the corrected document and the answer changes the same day. Fine-tuning earns its place when what needs to change is tone or a rigid output format, and it can sit alongside RAG, which still supplies the content and the citation.