GraphForge vs. LanceDB
When is vector-first retrieval enough, and when do relationships matter?
LanceDB is an embedded retrieval library on the Lance columnar format, with seven vector index types, hybrid BM25 and vector search with reranking, versioned tables, and writable branches. It does not model relationships. GraphForge does the reverse: the graph comes first, with exact vector and BM25 search alongside traversal and algorithms. Many retrieval pipelines need both. LanceDB handles approximate nearest-neighbor search at scale; GraphForge handles connections, paths, and communities.
Not a like-for-like swap. LanceDB is not a graph database. It is a vector-first alternative for retrieval workloads that might otherwise be built on a graph engine with vector search.
On this page
Should I choose LanceDB or GraphForge?
LanceDB for approximate vector search at scale; GraphForge when relationships shape the answer.
Choose GraphForge when
- Answers depend on how entities connect, through paths, neighborhoods, communities, and centrality.
- You want retrieval and graph traversal in one engine instead of two systems.
- You need provenance and recorded hypotheses for an investigation, not only dataset versions.
Choose LanceDB when
- Retrieval quality and scale depend on approximate vector indexes such as IVF-PQ or HNSW.
- Data is multimodal (images, embeddings, metadata) stored in a lakehouse format on object storage.
- You want versioned, branchable tables for datasets and embeddings.
What can each one do?
20 of 26 capabilities differ. For 11, LanceDB's docs did not confirm support, so they are marked Not established rather than No.
Each capability has a status and a source. “Not established” means we could not confirm the feature from LanceDB's primary documentation. It does not mean LanceDB lacks it.
| Capability | GraphForge | LanceDB |
|---|---|---|
| Storage | ||
| Property graph model | Shipped Labeled property graph with typed properties. 1 | Not established Not in LanceDB itself. lance-graph, a separate Cypher project in the lance-format organization, is outside LanceDB's docs. |
| Document model | Not offered No document-store model; properties accept nested, heterogeneous values. 1 | Partial Tables store images, video, audio, text and metadata together; no JSON document model. 17 |
| Relational tables | Not offered Not a relational database; results arrive as Arrow tables. 1 | Partial Tabular Arrow data with SQL filters; full SQL endpoint is Enterprise-only. 26 |
| Key-value access | Not offered No key-value API; data is a property graph. 1 | Not established No key-value API documented. |
| Columnar storage | Shipped Durable projects store graph data as Arrow and Parquet. 4 | Yes Lance is Arrow-native columnar storage optimized for random access as well as scans. 23 |
| Vector search | Shipped Exact cosine retrieval over embeddings you supply; no ANN index. 5 | Yes IVF_FLAT, IVF_PQ, IVF_SQ, IVF_RQ and IVF_HNSW_* indexes; l2, cosine, dot, hamming. 20 |
| Full-text search | Shipped Tantivy-backed BM25; hybrid text and vector retrieval through find(). 5 | Yes BM25 full-text search with phrase, boolean, boosted and fuzzy queries; hybrid search with reranking. 19 |
| Geospatial | Not offered No spatial types or indexes are documented. | Not established No geospatial types or indexes found in docs. |
| Execution | ||
| Transactional workloads (OLTP) | Partial Local analytical scope; durable atomic writes with queued and optimistic writer modes. 6 | Partial Row updates, deletes and upserts create versions; concurrent commits can hit retryable or incompatible conflicts. 2124 |
| Analytical workloads (OLAP) | Shipped Analytical verbs and Cypher over a local columnar project. 7 | Partial Columnar scans; analytical SQL via Enterprise Flight SQL endpoint (DataFusion). 26 |
| Graph traversal | Shipped Cypher MATCH patterns plus BFS, DFS, Dijkstra, A*, and reachability. 8 | Not established Traversal is not part of LanceDB's documented data model. |
| Variable-length paths | Shipped Bounded and unbounded patterns and shortestPath(). 8 | Not established Not documented in LanceDB or the lance-graph README. |
| Graph algorithms | Shipped 94 native algorithms behind analysis verbs, not Cypher CALL procedures. 37 | Not established No graph algorithms documented. |
| Incremental computation | Not offered Algorithms recompute on request; no incremental mode is documented. 7 | Not established No incremental view maintenance documented for OSS. |
| Deployment | ||
| Embedded / in-process | Shipped In-process library; no server, daemon, or connection string. 9 | Yes OSS LanceDB is an embedded retrieval library. 17 |
| Server mode | Planned Optional remote peer extension planned after 0.6.0. 11 | Paid edition Remote, cluster-backed serving of remote tables is LanceDB Enterprise. 27 |
| WebAssembly | Not planned A browser WebAssembly engine was declined in ADR 0027. 12 | Not established No official WebAssembly build documented. |
| Distributed cluster | Not planned Distributed clusters and HA replication are out of scope. 9 | Paid edition Enterprise separates query nodes, plan executors and indexers over shared object storage. 27 |
| Managed cloud | Not planned No hosted compute. The graphforge.sh Hub distributes repositories; execution stays local. 13 | Paid edition LanceDB Enterprise offers Managed (LanceDB-run accounts) and BYOC deployments. 25 |
| Investigation | ||
| Branching | Experimental Research versions, branches, slices, proposals, and forks ship in 0.6.0 release candidates. 14 | Yes Writable table branches in OSS and Enterprise; diff/merge APIs are Enterprise-only. 22 |
| Snapshots / time travel | Shipped Named checkpoints you can diff and revert, plus transaction-time snapshots. 3 | Yes Every write creates a table version; checkout older versions, restore, and tag them. 21 |
| Immutable history | Shipped Knowledge ledger records immutable assertions and evidence lineage. 16 | Partial Versions are retained until cleanup (default 7 days) unless tagged. 21 |
| Provenance | Shipped Provenance and knowledge ledger with evidence lineage. 16 | Not established No provenance feature documented. |
| Competing hypotheses | Shipped Immutable hypothesis groups with membership and selection events. 3 | Not established |
| Operations | ||
| ACID transactions | Partial Acknowledged-durable atomic publication; docs do not claim general ACID or serializable isolation. 6 | Yes Lance format spec: MVCC with ACID guarantees, optimistic concurrency, serializable version history. 24 |
| Replication | Not planned HA replication is out of scope. 9 | Not established No replication feature documented; Enterprise keeps durable data in object storage. |
Showing 26 of 26 capabilities.
How is each one built?
LanceDB: vector + tabular (lance format). GraphForge: an embedded property graph library in Rust, persisted as Parquet.
GraphForge
GraphForge is a Rust library that runs inside your process: no daemon, server, or connection string. Projects persist to a local directory as Arrow and Parquet data plus JSON metadata, and every query and analysis verb returns an Apache Arrow table. Algorithms run in the same native engine as Cypher and are called as API verbs (rank, cluster, paths, analyze, similar), not as Cypher procedures.
LanceDB
LanceDB is a Rust library over Lance, a columnar file format, table format, and catalog spec. Every write creates a table version committed atomically as a new manifest on the object store. Old versions are pruned after seven days by default unless tagged. Enterprise adds distributed query nodes, indexers, and a SQL endpoint.
Both are columnar and Arrow-native. LanceDB optimizes nearest-neighbor retrieval over large tables; GraphForge optimizes reasoning over relationships, with retrieval as one verb among several.
What the documentation says about LanceDB
- LanceDB OSS is described as an open-source embedded retrieval library with Python, TypeScript and Rust SDKs.17
- LanceDB is built on Lance, an open columnar lakehouse format with a file format, table format and catalog spec.23
- LanceDB supports seven vector index types, including IVF_PQ, IVF_RQ and IVF_HNSW_SQ.20
- Hybrid search combines vector and BM25 full-text search with reranking.19
- Every LanceDB write creates a new table version; restore makes an earlier version current without duplicating data.21
- LanceDB tables support isolated, writable branches forked from main or any other branch.22
- The Lance spec commits each transaction as a new manifest using atomic put-if-not-exists or rename-if-not-exists on the object store.24
- A separate lance-graph project in the lance-format organization provides a Cypher-capable graph query engine in Rust with Python bindings.28
Which one is faster?
Not measured yet. No head-to-head run has been published, so this page shows no performance chart.
Charts appear only after a reproducible run exists: pinned versions, disclosed hardware, generated data, raw results, and scripts.
No comparative results yet
The suites below apply to GraphForge and LanceDB. Each will be reported separately per benchmark class, never merged into one leaderboard.
Ingestion and writes
| Suite | Workloads | Metrics | Status |
|---|---|---|---|
| INGEST | CSV, JSONL, Parquet, graph edge lists | Records/s, MB/s, total elapsed | Not run |
Query, analytics, and system
| Suite | Workloads | Metrics | Status |
|---|---|---|---|
| LOOKUP | ID lookup, indexed property filter | p50/p95/p99 latency | Not run |
| HYBRID | Vector + graph + metadata filters | Recall@k, latency, throughput | Not run |
| VERSION | Branch, diff, merge, snapshot | Latency, storage overhead | Not run |
| EMBED | Startup, load, reopen, binary size | ms, MiB, disk footprint | Not run |
| SCALE | Increasing graph size and density | Throughput degradation, limits | Not run |
Which graph algorithms ship?
GraphForge: 94 native algorithms. LanceDB: no named algorithm library in its docs.
Algorithms are listed by name, as documented. Counts of extension modules or third-party libraries are not presented as native coverage.
GraphForge
94 Rust algorithms run in the engine behind five analysis verbs (rank, cluster, paths, analyze, similar) and return Arrow tables.37
A selection of the catalog:
- PageRank
- ArticleRank
- Betweenness (Brandes)
- Closeness
- Harmonic closeness
- Eigenvector
- HITS
- CELF influence maximization
- Triangle count
- K-core decomposition
- Louvain
- Leiden
- Label propagation
- Infomap
- Walktrap
- Girvan-Newman
- Weakly connected components
- Strongly connected components
- Dijkstra
- A*
- Bellman-Ford
- Yen's k-shortest paths
- Max flow / min cut
- Steiner tree
- Minimum spanning tree
- Topological sort
- Node2Vec
- GraphSAGE
- Fast random projection
- Node similarity (Jaccard)
LanceDB
LanceDB offers no graph analytics; its built-in retrieval is vector, full-text and hybrid search.19
Delivered as: n/a
No named algorithm catalog was found in the primary documentation.
Where can each one run?
GraphForge runs in-process only. LanceDB: embedded yes, server paid edition, cluster paid edition.
| Mode | GraphForge | LanceDB |
|---|---|---|
| Embedded / in-process | Shipped In-process library; no server, daemon, or connection string. 9 | Yes OSS LanceDB is an embedded retrieval library. 17 |
| Server mode | Planned Optional remote peer extension planned after 0.6.0. 11 | Paid edition Remote, cluster-backed serving of remote tables is LanceDB Enterprise. 27 |
| WebAssembly | Not planned A browser WebAssembly engine was declined in ADR 0027. 12 | Not established No official WebAssembly build documented. |
| Distributed cluster | Not planned Distributed clusters and HA replication are out of scope. 9 | Paid edition Enterprise separates query nodes, plan executors and indexers over shared object storage. 27 |
| Managed cloud | Not planned No hosted compute. The graphforge.sh Hub distributes repositories; execution stays local. 13 | Paid edition LanceDB Enterprise offers Managed (LanceDB-run accounts) and BYOC deployments. 25 |
| Replication | Not planned HA replication is out of scope. 9 | Not established No replication feature documented; Enterprise keeps durable data in object storage. |
What will my team write?
GraphForge: openCypher from Python, Node.js, Rust. LanceDB: SDK query builders (vector, full-text, hybrid), SQL filter expressions, SQL via Arrow Flight SQL (Enterprise only).
LanceDB is used through Python, TypeScript, and Rust query builders with SQL filter expressions. GraphForge uses Cypher plus a find() verb that combines BM25 and exact vector search over embeddings you supply.
- Query languages
- GraphForge openCypher
- LanceDB SDK query builders (vector, full-text, hybrid), SQL filter expressions, SQL via Arrow Flight SQL (Enterprise only)
- SDKs and bindings
- GraphForge Python, Node.js, Rust
- LanceDB Python, TypeScript, Rust, REST API (Enterprise)
- Written in
- GraphForge Rust
- LanceDB Rust
- Surfaces
- GraphForge Python, Node.js, Rust crates, CLI, VS Code extension, Notebooks, Agent skills
What are the trade-offs?
LanceDB does not model relationships. GraphForge's vector search is exact and needs your own embeddings.
A fair assessment cuts both ways.
GraphForge
- Vector search is exact (brute-force cosine), so it does not scale like approximate indexes.
- You must supply embeddings; GraphForge does not generate them.
- No object-storage-native table format.
LanceDB
- LanceDB is not a property graph database; relationships and traversals are not part of its documented data model.17
- Old versions are pruned by cleanup (7-day default retention), so time travel is bounded unless versions are tagged.21
- Under the Lance commit protocol some concurrent-write conflicts are retryable only by the application re-reading and rerunning, and some fail outright.24
- The SQL endpoint is Enterprise-only; OSS users query through SDK builders and SQL filter expressions.26
How would I migrate?
Load LanceDB data through Arrow, pandas, or Polars. Data access code is rewritten as Cypher.
Into GraphForge
- GraphForge 0.5.x loads data through Arrow, pandas, and Polars bulk calls or Cypher scripts; it has no native CSV loader.15
- Algorithms are API verbs (rank, cluster, paths, analyze, similar), not Cypher CALL procedures, so procedure calls become verb calls.7
- GraphForge passes the full openCypher 2024.3 TCK, so standard openCypher queries are the portable part of a migration.10
Moving from LanceDB
- Lance is Arrow-native, so node and edge tables can move to and from other Arrow-based tools without format conversion.23
- Relationships must be modeled as ordinary tables; lance-graph's GraphConfig maps Arrow tables to Cypher node labels by ID column, as in its README example.28
- Embeddings and retrieval queries do not map to Cypher; vector and hybrid search calls must be re-expressed in the target engine's vector index API.19
What else do people ask about LanceDB?
Can LanceDB store graphs? Not natively; use LanceDB for large-scale ANN and GraphForge for graph structure, side by side over Arrow.
Can LanceDB store graphs?
LanceDB's own docs say relationships and traversals are not part of its data model. A separate project, lance-graph, provides Cypher over Lance tables but is not part of LanceDB.
Should I use LanceDB or GraphForge for GraphRAG?
Use LanceDB when approximate vector search over large corpora is the bottleneck. Use GraphForge when the graph structure itself, such as paths and communities, shapes the context you retrieve. They can also run side by side, since both exchange Arrow.
Does GraphForge have approximate vector search?
No. GraphForge's find() runs exact cosine retrieval over the embeddings you supply, combined with BM25 full-text search.
How was this researched?
From 28 primary sources, accessed 2026-10-19, each cited where it is used.
Capabilities were checked against primary documentation, repositories, and release notes, accessed 2026-10-19. Each cell records two things: the capability’s state and the evidence behind it. Every GraphForge cell carries a release status. Shipped means stable 0.5.2; experimental means 0.6.0 release candidates only. GraphForge is pre-1.0 and makes no backward-compatibility guarantee yet.
- Yes documented natively
- Partial limited or caveated
- Extension official extension or module
- Paid edition paid or enterprise edition only
- Removed removed, with the version cited
- Not established not confirmed by our research; not a “no”
- Shipped in stable 0.5.2
- Partial shipped with documented limits
- Experimental in 0.6.0 release candidates only
- Planned on the roadmap
- Not offered outside the product's model
- Not planned explicitly out of scope
Benchmark classes
- Equivalent workloads. The same logical operation with the same correctness requirement on every engine.
- Native optimized workloads. Each engine's best supported implementation, including projections, indexes, extensions, and native algorithms.
- End-to-end application workloads. Time from raw data to a usable result: import, indexing, projection, computation, and output retrieval.
Sources · accessed 2026-10-19
- GraphForge repository README — GitHub
- GraphForge releases — GitHub
- GraphForge 0.5.0 changelog — GitHub
- Storage architecture — GraphForge Book
- Analytics integration: search and retrieval — GraphForge Guide
- ADR 0018: acknowledged durability and isolation — GitHub
- Algorithm verbs — GraphForge Book
- Cypher guide — GraphForge Guide
- openCypher compatibility — GraphForge Reference
- openCypher TCK compliance — GraphForge Reference
- Remote authority peer extension (#215) — GitHub
- ADR 0027: native runtime boundary — GitHub
- Release scope and roadmap — GraphForge Docs
- Keep research history — GraphForge Guide
- Graph construction — GraphForge Guide
- Knowledge public API v1 — GraphForge Book
- LanceDB documentation home — docs.lancedb.com
- Releases · lancedb/lancedb — GitHub
- Search — LanceDB docs
- Vector Search — LanceDB docs
- Versioning — LanceDB docs
- Branching — LanceDB docs
- Lance format — LanceDB docs
- Transaction specification — Lance format docs
- LanceDB Enterprise deployment guide — LanceDB docs
- Query with SQL (Enterprise-only) — LanceDB docs
- LanceDB Enterprise architecture — LanceDB docs
- lance-format/lance-graph README — GitHub
Try GraphForge on your own graph.
Install the package, open a project, and run the same notebook we measure with.