Home / Platform / Entity Graph
Platform · Records become knowledgeRetrieval finds text that looks like your question. It can't tell you that the plumber on this invoice is the same person in that text thread, who works on the property named in this lease, paid from the account on that statement. Answers to real questions live in the relationships between records — the part a pile of vector chunks throws away. Lossless keeps them. Every entity is a node, every relationship a sealed edge, the whole graph living in the same RLS-protected Postgres as your records.
The standard RAG stack chops your documents into 512-token slabs and embeds them. Ask a question, it finds the slabs whose vectors sit closest to your query's vector. That works for "summarize this PDF." It collapses the moment the answer requires hopping across two, three, five records that were never near each other in any document.
The chunk on the left and the graph on the right can hold the exact same words. Only one of them can answer "which of my tenants share a contractor?"
For two years the consensus was "embed everything, retrieve by similarity." In 2025–2026 that consensus broke. The investors, analysts, and infrastructure builders racing to fund the next layer of AI all landed on the same primitive — a knowledge graph as the memory substrate. Lossless has been building exactly that, per-person, since day one.
Every one of those voices is describing the enterprise version of this problem — a graph over a company's documents. Lossless is the first to build it where it actually compounds: a graph over a person's life, owned by that person, queryable by the agents they authorize.
Lossless ingests ~30 sources — Plaid, Gmail, Drive, Calendar, Airbnb, VRBO, Tesla, Coinbase, ChatGPT / Claude / Grok transcripts, and more. Every record is resolved into nodes and wired to the rest of your life by typed, provenance-backed edges. Hover any node to read the record and bloom its connections.
GraphRAG vendors build one shared graph over a company's documents. Lossless builds a sovereign graph per person — and that single architectural choice unlocks a stack of advantages that a multi-tenant enterprise graph can't replicate.
Your graph is not a partition of a shared database — it is its own entity graph in the same RLS-protected Postgres as your records. A query in one person's graph cannot, by construction, traverse into another's: the database engine refuses. Privacy is the topology, not a policy.
An enterprise graph asserts "A relates to B." Lossless asserts "A relates to B — here is the email, the statement line, and the parser version that proved it." Every edge is signed under the Universal Provenance Layer. Walk the graph, and you can walk the receipts.
The plumber's name on a Gmail receipt, his number in your contacts, his LLC on a Plaid transaction, his texts in iMessage — Lossless resolves all of them to one Person node. Without resolution, the graph is just disconnected aliases. Resolution is what makes it traversable.
Edges have an occurred-at and a known-at. The graph can answer "who was my contractor in March" and "what did I know about this account on the day I filed" — both. Amendments add edges; they never overwrite history. Discovery and audit become trivial.
Most "GraphRAG" is still vector search with a graph re-ranker stapled on. Lossless plans retrieval on the graph first — traverse the typed edges, then use embeddings to rank within the result set. Multi-hop questions get multi-hop answers.
The graph isn't locked inside one app. Claude, ChatGPT, Gemini, Grok — any agent can traverse it through MCP + REST, scoped to exactly the edges you permit. Your graph becomes the shared memory that every assistant queries and none of them owns.
Each of these is a multi-hop traversal — the answer is assembled by walking typed edges across domains. A vector index can't do it; a per-person entity graph does it in one query.
"Graph-first retrieval was the deciding factor — vector-only RAG keeps coughing up similar-looking text instead of the actual document we cited last week."
— Technical evaluator · litigation-support team
Models commoditize. A sealed, per-tenant graph compounds. See it work where a single-vector RAG would already be in trouble — a complex multi-party criminal-law matter.
Next pillar: Chronology. The graph plus a bi-temporal axis — every event has both an occurred-at and a known-at timestamp, so amended returns don't overwrite history and a charge that posts three days after the swipe doesn't break the timeline.