Each one reads and writes the same graph, under the same permission model. That is what makes them work together instead of stacking up.
Pick a module
Assistant
01 / 10
Ask a question, get an answer you can defend
The assistant reads your question in plain language, decides where to look, queries the graph and the documents, then writes. Every claim points to the document behind it: an answer is verified in one click rather than taken on trust.
Five search strategies, from pinpointing one entity to surveying a whole domain, selected automatically
Clickable citations: each source opens the entity record or the original document
Answers written as they come, with search stages shown live and kept in the history
Output that fits: text, table, diagram, or an Excel, Word, PowerPoint, PDF or CSV file to download
Private folders to file your conversations, project mode to share thread, files, memory and instructions with a team
Optional web search, enabled by the administrator, with the engine of your choice
Graph
02 / 10
The map of how your company connects
The graph is what separates Omniscient from document search. Your customers, contracts, projects, incidents and people are linked objects, and it is that structure which answers questions no single document contains.
Visual exploration: start from an entity, follow its links, filter by object or relationship type
Open model: add your own entity and relationship types without development work
Every node keeps its source documents, readable and downloadable from its record
Full-text search and advanced queries for technical users
Instant overview and smooth navigation on graphs with hundreds of thousands of nodes
A view of orphaned entities, to spot what hasn't been connected yet
Ingestion
03 / 10
Feed the graph without writing code
A pipeline is drawn: pick a source, describe what you want from it, schedule it. The engine extracts, structures, deduplicates and loads. Large files are read as a stream, without flooding the machine.
Connectors: dropped files and folders, structured files (CSV, TSV, Excel), SQL databases, REST APIs, Google Drive, SharePoint, HubSpot, Notion
Multimodal extraction: PDF, Word, PowerPoint, Excel and images, including text inside images
Entity and relationship recognition by language model, on your own business categories
Duplicate merging: exact, fuzzy and semantic matching, with alias handling
Assisted column mapping, result preview before the first run
Scheduling by hour, by day or by cron expression, and recovery after failure without re-ingesting everything
Connections
04 / 10
Credentials declared once, shared properly
A mailbox, a HubSpot account, a database: the connection is declared once, then referenced by the pipelines and automations that need it. A password that changes is corrected in one place.
Ten families: REST APIs, databases, Google Drive, SharePoint, HubSpot, Notion, outgoing and incoming mail, HTTP services, web search engines
The secret is never read back: no response ever returns a password, not even to the person who entered it
Explicit sharing: private, chosen people and roles, or the whole organisation with use-only rights
Rights checked twice, when saved then at every run: revoking a share really does stop the next job
Usage view: where a connection is referenced, before you change or delete it
Connection tests run by the service that will actually use it, so the verdict means something
Automations
05 / 10
What repeats no longer has to be done
An automation is a scenario you draw: a trigger, some steps, some effects. It reads your data, calls a language model, decides, writes to the graph, sends a message, and stops to ask your permission wherever you require it.
Over forty blocks: model calls, classification, structured extraction, summarising, translation, graph reads and writes, conditions, loops, aggregations, code
Effects: email, forwarding an original message with its attachments, Slack, Teams, in-app notification, HTTP request, outgoing webhook, Google Drive, Notion, document generation
Human approval: an approval step holds the scenario until someone decides, with a dedicated queue
Dry run: the scenario executes without producing real effects, showing what it would have sent
Execution log, dashboards, a register of side effects produced, and alerts when behaviour drifts
Template library: publish a proven scenario, instantiate it elsewhere, without carrying secrets across
Agents
06 / 10
Assistants that know one job well
An agent is an assistant narrowed to a domain, with its own instructions and data scope. A user's question is routed to whichever one can answer it, without them picking from a list.
Agents specialised by function, with their own instructions and restricted data access
Automatic routing based on the meaning of the question, falling back to generic handling when uncertain
Agent teams for processes that need several skills in sequence
Reusable skills: instruction blocks a user writes, shares, and each person switches on for themselves
Personal memory: preferences stated in conversation are remembered, reviewable and erasable at any time
Shared company context injected into conversations, so in-house vocabulary is understood
Predictive
07 / 10
Learning the shape of the graph to see what's coming
Graph models learn what your company normally looks like, then flag what departs from it. No data science skills are needed to put them to work.
Five graph neural network architectures, chosen to fit the task
Anomaly detection: the behaviour that doesn't resemble the others
Link prediction: the relationship that is likely but missing, such as the customer about to leave or the match worth making
Automatic classification and clustering of entities
Scheduled retraining, and alerts that trigger an automation directly
Notifications
08 / 10
Being told in the right place
Job finished, pipeline failed, approval requested, predictive alert: platform events become messages, on whichever channel each person picked.
Three channels: email, live in-app notification, signed webhook to your own systems
Customisable message templates, previewable before going live
Preferences per person, per channel and per event type
Automatic retry on delivery failure, with increasing intervals
Delivery history and manual resend
Monitoring
09 / 10
Knowing what the platform does, and what it costs
Every job leaves a usable trace: what ran, what failed and why, how many tokens were spent and for how much. The AI bill stops being an end-of-month surprise.
Pipeline console: runs from every origin, searchable logs, failures grouped by cause
Cost accounting: generation tokens and embeddings counted separately, per job and per period
Health alerts raised by the platform itself, without depending on an external tool
Service status, inventory of deployed versions, CSV or JSON export of whatever is on screen
Governance
10 / 10
Permissions first, search second
A platform that reads the whole document estate is only worth having if it honours existing access rules. Filtering happens during the search: an answer cannot rest on a document its reader may not open.
Five predefined roles plus custom roles, by resource and by action
Fine-grained access rules on entities, relationships and documents
Sharing decided object by object: private, named people and roles, or the whole organisation
Strict separation between organisations in shared mode, including in the graph and the indexes
GDPR-compliant audit log, exportable for your reviews
Encryption in transit, and secrets encrypted at rest
Frequently asked questions
Which sources can the platform connect to?
Document spaces first: SharePoint, Google Drive, Notion, and plain file upload when a source has no API. Then business applications, with HubSpot for customer relationships and any REST API for the rest. Finally databases, PostgreSQL, MySQL, SQL Server and MongoDB, as well as mailboxes over IMAP. Each connection is declared once, shared with whoever needs it, and reused by ingestion pipelines and automations alike.
Which document formats can it read?
The usual office formats, PDF, Word, Excel and PowerPoint, tabular files as CSV, and images, whose text is extracted by optical character recognition. A scanned PDF is therefore read like a born-digital one. Spreadsheet tables are handled row by row rather than as a block, which makes it possible to retrieve one precise row from a file holding tens of thousands of entries.
How is this different from a document management system or an enterprise search engine?
Document management stores files and hands back the one you name. Enterprise search goes further and finds a file from words it contains. Neither answers a question whose answer is written in no single file, because it follows from several. That is what the knowledge graph does: it links objects to one another, and semantic search works on the meaning of your question rather than on its words.
Do we have to replace our existing tools?
No. Omniscient reads your sources where they already are and takes the place of none of them. Your documents stay in your document space, your customer data in your customer tool, your tables in your databases. The platform builds the map that links them and answers on top of it. That is also what makes leaving simple: what you would lose by stopping is the map, not the data.
Is this only for large companies?
No, and the most common need is in fact the opposite. A small or mid-sized company has no less scattered data than a large group, it has fewer people available to go and find it. A first scope runs on two or three sources, in shared mode, with nothing installed on your side. Dedicated and on-premise modes exist for organisations whose data cannot leave the network.
Which generative AI model is used, and can we choose it?
Yes. The platform is tied to no language model vendor: you can use ours, your own, or an open model running on your own infrastructure. In that last case no data leaves your network, generation included. The model only understands the question and writes the answer. What it writes from always comes from your data, and that does not change with the model you pick.