The self-assembling graph database
Your documents change. Your team's questions change. Spyndel is a graph database that writes itself from your documents and keeps up: it finds evidence across documents, flags what a revision may have made out of date, and learns from reviewers' corrections to retrieve better next time.
No LLM makes the connections. Spyndel's own compact models find and check every one, on your hardware: no per-document LLM extraction charges, however large your library grows.
One loop, running on its own
- IngestDocuments become individual statements
- OrganizeStatements find their topics and categories
- ConnectRelated ideas are linked across documents, by Spyndel's own models: no LLM, no tokens
- RetrieveAI answers draw on statements and their connections
- AdaptWhat answers use, and what reviewers correct, shapes the next answer
Spyndel builds itself
Give it documents. Spyndel finds the statements in them, discovers the subjects they cover, and connects related ideas across documents. You never design a schema, tag anything or write extraction rules: the graph database fills itself.
Self-constructing
Statements, topics, categories and the connections between them are built for you, from the documents themselves.
Verified connections
Links are made only when two statements are actually about the same thing, not just worded alike, and they are checked by Spyndel's own models, not a large language model. Every link carries a calibrated confidence score.
Statement-level precision
Search and links work on individual statements, and every statement keeps its document, page and section, with the text around it.
Connections without the LLM extraction bill
LLM-built knowledge graphs send every page of every document to a large language model to work out what's connected, and charge you for every token. Spyndel doesn't use an LLM to make connections at all: the only cost is your own GPU or CPU time. Plain vector search spends nothing on connections too, because it makes none.
in LLM tokens to find and check connections
less AI spend than the LLM-extraction graph tools we tested
to check 13,393 candidate connections on one GPU
Measured on 100 public NRC documents, October 2026, with Claude Sonnet 5 at AWS list price for every system that used an AI model. Spyndel's whole build spent $0.68, all of it on optional topic names and document summaries; none on connections.
Spyndel grows as knowledge changes
Add a document and it connects into everything already there, not onto the end of a list. As the collection grows, Spyndel reorganizes it.
- Connected on arrival. A new document's statements link to related statements across the whole library, and older statements gain connections to it.
- Reorganized as it grows. Rebuilding finds the topics again across everything, so new subjects get topics of their own.
- Cross-document insight. See where documents agree and overlap, and which ones cover the same requirement.
for a 400-page document to join the graph on one GPU
of statements in a single graph, tested on large engineering and regulatory libraries
Spyndel learns from use, and from you
When AI answers rely on a connection, that connection ranks higher next time. And when search misses something, one correction from a reviewer fixes the next search and the ones after it, in your team's own words.
9% → 44% of previously failed questions answered completely after a reviewer supplied the missing evidence once, measured on 112 new wordings of those questions; answers scored +0.24 higher, graded blind, 28 better and 6 worse. +4.2 points more of the right guidance found after 240 answered questions, with no question losing the document it needed. See the measurements
Answers with sources
Ask questions in plain language. Every answer cites the exact statements it relies on, and every citation is checked before you see it.
Plugs into your AI
A built-in Model Context Protocol (MCP) endpoint lets Claude, ChatGPT and other assistants search, walk the graph and pull full source attribution.
Sharper on familiar needs
Each answered question records which statements and connections it used, so the questions your team keeps coming back to are answered better over time.
Fix a miss once
When search misses the statement an answer needed, a reviewer points to it. From then on that evidence comes back for the same question however it is worded. The decision keeps who made it and why, and reverts in one click; the graph itself is never rewritten.
Your words, its words
Your people say "backup generator"; the documents say "emergency diesel generator". Spyndel notices the mismatch in the corrections, proposes the pair, and once a reviewer confirms it, every search looks for both.
Change alerts
When a revised document arrives, Spyndel lists what changed and which statements elsewhere repeat or depend on it, so a revision becomes a review list instead of a surprise.
A graph database, without the database work
Spyndel is a graph database for documents. What sets it apart is who writes the graph.
Conventional graph database
You design the schema, write the loaders and keep the graph current. The right tool for data that is already structured; a project of its own for a library of documents.
LLM-built knowledge graph
A large language model reads every page to extract entities and relationships. Automatic, but billed per page, every time, and your documents leave your network to be read.
Spyndel
Reads the documents itself, discovers the structure, and verifies every connection on your own hardware. No schema, no loaders, no token bill for connections. Still a graph you can query.
What is a self-assembling graph database? Read the explainer →
Start building your graph today
The Community Edition is free to download and runs on your own machine. Need scale, full impact analysis, commercial support or a managed deployment? See Professional.