Taxonomy first
Your own categories and terms define the graph.
Early accessDemo coming soon
Every graph can be represented as a matrix.
The taxonomy-first graph database. Your records become a knowledge graph, with ranked, explained matches in milliseconds.
- Taxonomy first
- ETL built in
- No LLM calls
- No data copy
- Every match explained
- Built for agents
- 244×faster than Neo4j Community
- 1.9×faster than Elasticsearch
- 17 sto write 1M records
- 1.0 GBmemory for 1M records
Typical search, the same 1M records and limits for every engine. How we tested
AI models and the capabilities that describe them form a graph.
Speed
Same searches. Same answers. Milliseconds, not seconds.
2.9 ms for a typical search at 1M records
The median: half of all searches finish faster.
- Matrie2.87 ms
- Elasticsearch5.36 ms1.9×
- Neo4j Community701 ms244×
- FalkorDB2,493 ms868×
All four engines ran the same searches on the same 1M records and returned their top 10. Bars are to scale. See every result.
Why Matrie
Your knowledge graph, at matrix speed
- Your database
- ETL scriptsYou build them
- LLM extractionA bill per record
- Graph or vector DBA second copy to sync
- Your app
- Your database
- MatrieExtracts, indexes and ranks. No LLM, no copy
- Your app
Hierarchical buckets
Nest terms as deep as you like, with weights and must-haves per group.
ETL built in
Connect, extract, clean and stay in sync. No pipeline to build.
Auto Enhancer
Finds the phrases your taxonomy is missing, and rates each one.
No LLM calls
No model bill per search, and the same input always ranks the same.
No data copy
Records stay in your database. The index is 164 MB for 1M records.
Every match explained
What matched, what's missing, the score and the distance.
Built for agents
12 tools for AI agents over MCP, plus a JSON API for apps.
- Also included
- Must-have rules
- Location radius
- Matching in both directions
- Background syncs
- Web console with user roles
How it works
From your records to a ranked answer
-
Your records
-
Your taxonomy
Free text becomes the terms you defined
-
The matrix
Each term a record has becomes one lit cell
-
Ranked matches
Filtered, weighted, ordered and explained
1M records, from source database to searchable, in about 8.5 minutes. Projected from a measured run of the full pipeline: 358,452 records in 183 seconds.
Agents
A matching tool your agents can trust
- Answers the same way every time
- Gives reasons your agent can quote
- Returns short lists, so prompts stay small
userFind an open-weight coding model we can host in the EU.
toolsearch("code generation, open weights, EU hosting")
- M-1042code generation, open weights, EU3 of 3
- M-0771code generation, open weights, EU3 of 3
- M-2310open weights, EU2 of 3
agentM-1042 is the best fit. It matches all three terms and ranks first.
Use cases
Built for the large datasets you already have
Large record sets, and clear rules for what counts as a match.
Who it's for
AI and agent teams
Facts your agents can quote, with no model bill per lookup.
Data platform teams
A knowledge graph over the database you already run.
Marketplaces
Match each request to providers with the right skills, licence and distance.
Regulated data owners
A reason for every result, and source text that stays with you.
Search and ML teams
A fast rule-based first pass before your re-ranker or LLM.
Search
Start from
Made-up data, real logic.
Also a good fit Each starts from a taxonomy that already exists
- Insurance claimsDiagnosis and procedure codesICD-10, CPT
- Product catalogsStandard product categoriesGS1 GPC
- Patents and researchPatent classes and medical subjectsCPC, MeSH
- Threat intelligenceAttacker tactics and techniquesMITRE ATT&CK
- Provider directoriesHealthcare provider specialtiesNUCC
- Legal and complianceYour approved contract clausesYour clause library
Choose something else for
- Multi-hop analysis, such as fraud rings
- Similarity search with no taxonomy
Benchmarks
1M records, four engines, every answer checked
- Same data1,000,000 records, 12.47M links
- Same searches957, in the same order
- Same processor4 cores, 8 threads
- Same memory limit4.5 GB each
- One worker eachOne search at a time
- Same answerTop 10: 3,146 checked, none wrong
| EngineMedian time per searchBrighter is betterBest in its column↓ lower or ↑ higher is better | Shortms↓, lower is betterAbout 4 keywords | Typicalms↓, lower is betterA real record's 6 keywords | With radiusms↓, lower is betterTypical, within 50 km | With required termms↓, lower is betterTypical, plus a must-have | Broadms↓, lower is better6 common keywords | Searches/sper s↑, higher is betterA mix of all five |
|---|---|---|---|---|---|---|
| Matrie | 1.12 | 2.87 | 1.27 | 1.25 | 32.2 | 93.4 |
| Elasticsearch | 4.95 | 5.36 | 5.63 | 8.15 | 39.7 | 56.8 |
| Neo4j Community | 44.6 | 701 | 849 | 23.7 | 1,183 | 1.62 |
| FalkorDB | 56.0 | 2,493 | 2,942 | 18.6 | 3,287 | 0.42 |
| Coming soon1B-scale benchmarks | ||||||
Hover a cell for the slower searches.
Where Matrie wins
- Fastest median on all five search types, by 10% or more
- Most searches a second: 93.4
- Fastest to write 1M records into a running database: 17 s
- Quickest to take a new record and return its 10 most similar: 23.9 ms
- Least memory: 1.0 GB at peak
Where it's close
- Nearest rival on a typical search: Elasticsearch, 1.9× slower
Where Matrie loses
- No vector search. Similar records are found through the taxonomy terms they share, and Auto Enhancer suggests the terms your taxonomy is missing
- Elasticsearch's index is smaller on disk: 119 MB to Matrie's 164 MB
Writing 1M records time↓, lower is better
Into a running database, 1,000 per request.
New record, then its 10 most similar ms↓, lower is better
What an app waits for with each new record.
Memory while searching GB↓, lower is better
Peak, with the typical level marked, under the same 4.5 GB limit.
Saving writes: Neo4j and Elasticsearch save each one to disk before confirming it. Matrie saves on demand, and its time includes the final save. FalkorDB kept writes in memory only, with saving off as in the search tests. Neo4j's offline importer loaded the same 1M records in 19 s, but it only works on an empty, stopped database.
Evidence: every number, and how we got it
- 957 searches built from real records, in 5 types
- 4 engines, one at a time, in the same order
- Top 10 scores compared with the known answer
- 3,146 answers checked, none wrong
Setup
- When
- 3 November 2026, one laptop
- Machine
- Ryzen 9 5900HX, 16 threads, 31 GB, other programs running
- Per engine
- 8 threads, 4.5 GB, 1 worker
- Data
- 2 copies of our 358,452 real records, plus 283,096 sampled
- Source
- 9,876 keywords, 4,468,219 links. Private, so not published
Fair to every engine
- A warm-up before timing
- The same keyword weights for all
- A hand-written query plan for Neo4j and FalkorDB
- Elasticsearch merged to one segment, its fastest
- 2.5 GB Java heap for Elasticsearch and Neo4j
Before you quote it
- A speed test, not a test of match quality
- A synthetic mix of real records
- The full pipeline time, about 8.5 minutes, is projected
- Matrie's default tie-break sort, off here, makes searches 1.5 to 1.9 times slower
| Engine | Version | How it ran | Searches runcount | Answers checkedcount | Wrongcount↓, lower is better |
|---|---|---|---|---|---|
| Matrie | Development build | Service on a saved snapshot | 957 | 957 | 0 |
| Elasticsearch | 8.19.0 | Server in Docker | 957 | 957 | 0 |
| Neo4j Community | 5.26.30 Community | Server in Docker | 838 | 838 | 0 |
| FalkorDB | 4.22.3 | Server in Docker | 394 | 394 | 0 |
| Engine | Shortms↓, lower is better | Typicalms↓, lower is better | With radiusms↓, lower is better | With required termms↓, lower is better | Broadms↓, lower is better |
|---|---|---|---|---|---|
| Matrie | 6.19 | 52.5 | 3.94 | 18.0 | 46.6 |
| Elasticsearch | 15.1 | 102 | 17.7 | 72.4 | 60.4 |
| Neo4j Community | 1,009 | 1,588 | 1,775 | 2,486 | 1,384 |
| FalkorDB | 3,400 | 3,586 | 3,935 | 5,091 | 3,578 |
95 of every 100 searches finished within this time.
| Engine | Searches a secondper s↑, higher is better | Searches in one minuteper min↑, higher is better |
|---|---|---|
| Matrie | 93.4 | 5,604 |
| Elasticsearch | 56.8 | 3,409 |
| Neo4j Community | 1.62 | 97 |
| FalkorDB | 0.42 | 25 |
| Engine | Write 1M, runningtime↓, lower is better | One at a timeper s↑, higher is better | New + similarms↓, lower is better | Initial bulk loads↓, lower is better | Peak memoryMB↓, lower is better | DiskMB↓, lower is better |
|---|---|---|---|---|---|---|
| Matrie | 17 s | 2,223 | 23.9 | 20 | 996 | 164 |
| Elasticsearch | 33 s | 319 | 45.8 | 27 | 3,149 | 119 |
| FalkorDB | 109 s | 1,029 | 5,841 | 26 | 2,892 | RAM only |
| Neo4j Community | 2.6 min | 447 | 3,014 | 19 | 3,473 | 1,071 |
"Write 1M, running" and "New + similar" are the charts above. "One at a time" is one record per request, for a minute. The bulk load uses each engine's own bulk tool. None of these includes reading free text, which the full pipeline does. Memory was read every 2 seconds.
Compare
Head to head, in numbers
Same test, same 1M records, same limits. A tick marks the better result, ↓ means lower is better and ↑ higher, and bars are to scale.
Matrie vs Neo4j Community
244× faster on a typical search
- Typical search2.87 ms701 ms
- Broad search32.2 ms1,183 ms
- Searches a second93.41.62
- Write 1M records17 s2.6 min
- New record + similar23.9 ms3,014 ms
- Peak memory1.0 GB3.4 GB
- Disk164 MB1,071 MB
Choose Neo4j for multi-hop queries, Cypher and graph algorithms.
Matrie vs FalkorDB
868× faster on a typical search
- Typical search2.87 ms2,493 ms
- Broad search32.2 ms3,287 ms
- Searches a second93.40.42
- Write 1M records17 s109 s
- New record + similar23.9 ms5,841 ms
- Peak memory1.0 GB2.8 GB
Choose FalkorDB for multi-hop traversal with Cypher, and GraphRAG tooling.
Matrie vs Elasticsearch
1.9× faster on a typical search
- Typical search2.87 ms5.36 ms
- Broad search32.2 ms39.7 ms
- Searches a second93.456.8
- Write 1M records17 s33 s
- New record + similar23.9 ms45.8 ms
- Peak memory1.0 GB3.1 GB
- Disk164 MB119 MB
Choose Elasticsearch for full-text relevance, vector search and scale-out today.
Run this test on your own data
The same engines and conditions, on your records, as part of a pilot.
Feature by feature
| Capability | Matrie | Neo4j | FalkorDB | Elasticsearch |
|---|---|---|---|---|
| Best suited for | One-hop matching and ranking | Multi-hop graph queries | Fast graph traversal | Full-text search and analytics |
| Records stay in your database | An index only | Copied into Neo4j | Copied, in memory | Copied into indices |
| Text to facts, with no LLM | Built in, from your taxonomy | Yours to build | Yours to build | Text analysis; the mapping is yours |
| Pipeline and sync built in | Connectors, extraction, sync | Import tools, LOAD CSV | CSV bulk loader | Connectors, ingest pipelines |
| Must-have rules, no code | Set in configuration | Written into each query | Written into each query | Filters in each query |
| Weighted ranking, no code | Weights per group of terms | Coded into each query | Coded into each query | Scoring in each query |
| Location radius | Built in | point.distance | distance() | geo_distance |
| Graph traversal | One hop, plus two-hop similarity | Any depth | Any depth | None |
| Query language | openCypher, coming soon. API and MCP tools today | Cypher | A subset of openCypher | Query DSL, ES|QL |
| Vector search | Not built in | Vector indexes | Vector indexes | kNN on dense vectors |
| MCP for agents | Built in, 12 tools | Official server | Available | Agent Builder server |
| Durability and scale-out | Snapshots; one machine today | ACID; clusters in Enterprise | RDB and AOF, replication | Shards and replicas |
From vendor documentation read on 2 November 2026, in our words: Neo4j, FalkorDB, Elasticsearch.
Pricing
Contact us for pricing
Pricing depends on your workload. Tell us how many records you have and how many searches you expect, and we'll send you a quote.
Ask for a price
FAQ
Questions a careful buyer asks
What does "taxonomy-first" mean?
You describe what matters in your domain: its categories, its terms, other words for those terms, and which ones are must-haves. Matrie builds the graph from that description, so results follow rules you can read and change.
Is Matrie really a graph database?
Yes, with a narrow focus. It holds a knowledge graph of your records and the terms that describe them, stored as a matrix. It answers one-hop matches and two-hop similarity. It doesn't run arbitrary paths or graph algorithms, so if you need those, choose Neo4j or FalkorDB. Support for openCypher, the open graph query language, is coming soon.
Does it use an LLM?
Not when data loads or when someone searches. The optional Auto Enhancer uses a small model on your own hardware to rate the key phrases it suggests.
Can it work with RAG and AI agents?
Yes, as the retrieval step. Rules narrow millions of documents to the few worth sending to your model, and agents call it over MCP. Matrie finds and ranks. It doesn't write answers, and it has no vector search.
How is this different from "no ETL" graph tools?
They run graph queries over tables you've already structured. Matrie also reads the free text in your records and maps it to your taxonomy. In both cases the source data stays where it is.
Where does my data live?
Source text stays in your databases. Matrie reads it to build the index and when someone opens a record, and doesn't keep it. The index holds ids, coded keywords and coordinates, and connection passwords are stored encrypted. Coded keywords can still be sensitive, so protect the index as you would any internal data store.
Which data sources does the ETL support?
MongoDB and PostgreSQL today, with support for other databases on the way. Tell us what you use, and we'll tell you when it's ready.
What does "1M records" mean here?
1,000,000 records built from our real dataset of 358,452. Every engine handles that size on one machine, so the comparison is about the engines rather than how much memory each was given. 1B-scale benchmarks are coming soon. The evidence section shows how we built them.
What happens if the service crashes?
Matrie saves snapshots, which are saved copies of the index, so a crash loses changes made since the last one. A re-sync restores them from your databases, which stay the source of truth. Today Matrie runs on a single machine.
When is the demo?
Soon. Leave your details and we'll let you know when it opens. We can also run the benchmark on your own data.
Demo coming soon
Be first to try Matrie
Leave your details and we'll tell you when the demo opens. Ask about pricing or a pilot on your own data any time.