Enterprise search and RAG
One question, answered from everything your organisation holds, under the permissions of the person asking.
The knowledge exists. Nobody can reach it. #
Organisational memory is scattered across document stores, mailboxes, shared drives, databases and a dozen SaaS tools. People know the answer is somewhere; finding it is a research project.
Retrieval bolted onto that estate tends to fail in two directions at once. It misses what it should find, because keyword search cannot follow meaning and semantic search cannot match an exact part number. And it surfaces what the reader was never cleared to see, because the index forgot the permissions of the system it copied from.
An answer a person is not cleared to receive is not retrieved in the first place.
Where it breaks
- Keyword search misses meaning; vector search misses identifiers, codes and exact terms.
- Indexes are built by copying documents, and the copy loses the source system's access rules.
- Answers arrive with no citation, so a reader cannot check them and will not trust them.
- Every new source is another connector, another index and another thing to keep fresh.
One lakehouse, hybrid retrieval, permissions enforced in the query #
Atom holds your material as one per-tenant lakehouse: uploaded documents, connected systems, structured tables and external databases federated in place. Retrieval runs over that single surface, fusing keyword precision with semantic recall into one ranked result.
Permission filtering is part of query execution rather than a filter in front of it. Row filters, column masks and deny decisions are evaluated on every read through one shared path, so the same rules apply whether the question comes from the interface, an application, an agent or a database client. An answer a person is not cleared to receive is not retrieved in the first place.
- Atom
- The lakehouse, hybrid search, ingestion pipelines, federation and the governance path every read passes through.
- Anima
- The embedding models that make retrieval semantic and the language model that composes the answer, served privately.
- Platform Foundation
- Identity, the authorisation decision behind every filter, and the audit ledger that records each access.
What the retrieval path does #
Grounded answers depend on unglamorous things: getting the content in, indexing it two ways, and never losing the permissions attached to it.
Hybrid ranking, tunable
BM25 keyword search and vector semantic search are fused with reciprocal-rank fusion into one relevance-ranked result, tunable from pure keyword to pure semantic per query.
Sync from the systems you already run
Connectors incrementally bring files and mail across from Microsoft SharePoint and OneDrive, Google Drive, Dropbox, S3 and Gmail or IMAP, per tenant, without a bulk migration project.
Documents become retrievable structure
Layout, tables, figures and OCR are extracted from PDFs, Office formats, spreadsheets and images, then split into overlapping, de-duplicated segments sized for retrieval.
Both indexes over the same data
Inverted indexes handle exact terms and identifiers; vector indexes handle meaning. Both are built over the same open columnar datasets rather than a separate copy.
Fresh as of the last write
Every change publishes to a live change feed, so indexes, live queries and downstream consumers follow writes in real time instead of on a nightly rebuild.
Answers that show their working
Retrieved passages are cited back to the exact source, and generated analytical queries can be inspected alongside the answer.
External databases without a copy
External PostgreSQL, MySQL and ODBC systems register as native tables and work is pushed down to them, so a single question can span the lakehouse and systems of record.
A record of every access
A canonical, redacted audit ledger records each data access correlated to the identity and query behind it, which is what makes a retrieval system reviewable.
Where the corpus can stay #
Retrieval runs where the documents are allowed to be, including estates with no external route at all.
- Managed sovereign cloud
- Fastest route to a working corpus, still inside your boundary.
- Private cloud
- Alongside the systems the connectors read from.
- On-premises
- For content that is not permitted to leave the building.
- Air-gapped
- Embeddings and generation served entirely from local weights.
Try it against your own corpus
Bring a body of material and the access rules that govern it. The interesting test is not whether retrieval finds things, but whether it refuses to find the things it should not.