Skip to content

Releases

v0.2.1

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 14 Nov 18:18
pip install -U selfgraph

New

  • known_at on retrieve and check_answerable reads only what the graph had stored by that time, so a backdated correction does not rewrite past reads. CLI --known-at, MCP known_at.
  • Relation.certifications keeps every certification with by, at, and until.

Changed

  • Certifying again by the same person, or uncertifying an uncertified relation, changes nothing.
  • Inserting the same text again keeps the source version as first stored.

v0.2.0

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 14 Nov 02:55
168d6c6
pip install -U selfgraph

New

  • SelfGraph(level="light" | "balanced" | "max") sets how much each capability does, from deterministic to LLM-backed. max_tier caps cost.
  • check_answerable(question) says whether the graph holds what a question asks about. readiness() checks the store and lints the data. No LLM.
  • relations.certify(id, by=) marks a fact as checked by a person, so contradictions, source replacement, prune and dedupe keep it.
  • insert.text(..., replace_source=True) retires facts a new source version no longer states.
  • contradiction="llm" retires conflicting facts without a schema.
  • Schema types take synonyms, definitions and mappings. Schema.from_tables drafts a schema from table metadata.
  • retrieve(..., skip_hubs=True) stops walking through entities linked to everything.
  • schema.to_ossie() exports an Apache Ossie ontology. export_obsidian() writes an Obsidian vault.
  • HydraDB writes run in parallel, with indexed lookups.
  • New CLI commands: readiness, check-answerable, relation certify, export obsidian. New MCP tools: check_readiness, check_answerable, certify_relation, uncertify_relation.

Breaking

  • Renamed: Ontology to Extraction, DefaultExtractor to LLMExtractor, BasicExtractor to NERExtractor, preference kind to stance, and maintenance.consolidate to maintenance.dedupe_relations.

ts-v0.2.1

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 14 Nov 18:18
npm install @hydradb/selfgraph

New

  • known_at option on retrieve and checkAnswerable reads only what the graph had stored by that time, so a backdated correction does not rewrite past reads. CLI --known-at, MCP known_at.
  • Relation.certifications keeps every certification with by, at, and until.

Changed

  • Certifying again by the same person, or uncertifying an uncertified relation, changes nothing.
  • Inserting the same text again keeps the source version as first stored.
  • Relation has two new required fields, invalidated_known_at and certifications.

ts-v0.2.0

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 14 Nov 02:55
168d6c6
npm install @hydradb/selfgraph@0.2.0

New

  • new SelfGraph(store, { level: "light" | "balanced" | "max" }) sets how much each capability does. max_tier caps cost.
  • graph.checkAnswerable(question) and graph.readiness() tell an agent what the graph holds. No LLM.
  • graph.relations.certify(id, { by }) marks a fact as checked by a person, so contradictions, source replacement, prune and dedupe keep it.
  • insert.text(text, { replace_source: true }) retires facts a new source version no longer states.
  • contradiction: "llm" retires conflicting facts without a schema.
  • Schema types take synonyms, definitions and mappings. fromTables(tables) drafts a schema from table metadata.
  • retrieve(query, { skip_hubs: true }) stops walking through hub entities.
  • toOssie(schema) exports an Apache Ossie ontology. graph.exportObsidian() writes an Obsidian vault.
  • HydraDB writes run in parallel, with indexed lookups.
  • Same CLI commands and MCP tools as Python, and the same output byte for byte.

Breaking

  • Renamed: Ontology to Extraction, DefaultExtractor to LLMExtractor, BasicExtractor to NERExtractor, preference kind to stance, and maintenance.consolidate to maintenance.dedupeRelations.

v0.1.0

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 09 Nov 18:28
pip install -U selfgraph
  • Entity resolution: fold a new entity into a stored one that names the same thing. Modes are exact (default), default (shared external id or alias), fuzzy (similar names) and llm (an LLM confirms close pairs). Every merge is recorded and can be undone with unmerge. Available as --resolver in the CLI and as entity_merges and unmerge_entity in MCP.
  • Contradictions: a schema can cap relations per entity with max_count, e.g. one employer per person. A newer fact that breaks the cap warns by default. In supersede mode the older fact is retired with invalidated_at and stays in history. relation list --include-invalidated shows retired relations.
  • as_of reads: retrieve(as_of=...) and --as-of read the graph as it was at a given time.
  • Maintenance: decay scores items by recency and frequency, prune removes weak or old items, and consolidate merges duplicate parallel relations. All three only show the plan unless you pass --apply (dry_run=False in the SDK).
  • Neo4j schema interop: selfgraph schema import and selfgraph schema export convert to and from a Neo4j graph schema.
  • Evidence: every entity and relation records the source text that supports it (entity evidence, relation evidence).
  • Schema checks now also apply to single create and update calls. The schema also supports min_count/max_count, parent types and required properties.
  • HydraDB 0.4: writes are batched.
  • CLI: versions subcommands, retrieve --context-only, connection flags after the subcommand, inline JSON for --schema, progress on stderr, and settings checked before connecting.
  • Fixes: HydraDB now clears emptied fields, re-inserting an old fact no longer retires a newer one, alias-form ids map to the real entity, and an invalid LLM key or model now fails with an error instead of silently producing an empty graph.

Breaking changes

  • The llm extra no longer installs mcp. Use pip install "selfgraph[mcp]" for the MCP server.
  • List cursors are now id-based. Cursors saved from 0.0.x no longer work.
  • A time without a timezone is read as UTC.
  • CLI exit codes: 2 for invalid input, 1 for runtime errors, 130 when interrupted.

Docs: https://github.com/hydra-db/selfgraph#readme

ts-v0.1.0

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 09 Nov 18:28
npm install @hydradb/selfgraph@0.1.0

Same changes as Python 0.1.0: entity resolution with merge and unmerge, contradictions with supersede, as_of reads, maintenance (decay, prune, consolidate), Neo4j schema interop, evidence records, schema checks on single writes, HydraDB 0.4 batching, and the CLI and MCP additions and fixes. Also:

  • setLogger routes the library's warnings and progress lines to your own logger.

Breaking changes

  • The ai peer dependency is now >=6.
  • List cursors are now id-based. Cursors saved from 0.0.x no longer work.
  • A time without a timezone is read as UTC.
  • CLI exit codes: 2 for invalid input, 1 for runtime errors, 130 when interrupted.

Docs: https://github.com/hydra-db/selfgraph/tree/main/typescript#readme

v0.0.4

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 07 Nov 03:21
pip install -U selfgraph
  • Optional schema: declare entity and relation types to guide extraction and check inserts, in warn, strict or off mode. Also available in the CLI (--schema, --schema-mode, selfgraph schema) and in the MCP server (get_schema tool).
  • Clear errors for a wrong password, an unreachable server or a malformed URI.
  • Empty or blank names and ids now raise ValidationError.
  • Kuzu creates missing parent folders. memgraph is accepted as a store name.
  • New CLI flags for external ids, removing aliases and properties, event times, and preference versions.
  • SyncSelfGraph.connect() and SyncSelfGraph.from_env().
  • A prompts_ignored warning when prompts are set but no LLM is configured.
  • Docs fixes.

Docs: https://github.com/hydra-db/selfgraph#readme

v0.0.3

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 07 Nov 01:31

First public release of SelfGraph: turn text into a queryable, versioned knowledge graph, with one API over any graph backend.

pip install selfgraph
  • Extract entities, relations, preferences and decisions from text with any LLM, or without one via spaCy or a built-in fallback.
  • Typed CRUD with deterministic ids, versioned relations, source provenance and namespaces.
  • Retrieval by keyword seeds plus BFS, DFS or personalized PageRank. No vector store.
  • Backends: Neo4j, Memgraph, FalkorDB, embedded Kuzu and HydraDB.
  • A CLI and an MCP server over the same API.
  • A TypeScript port on npm as @hydradb/selfgraph.

Docs: https://github.com/hydra-db/selfgraph#readme

ts-v0.0.4

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 07 Nov 03:22
npm install @hydradb/selfgraph@0.0.4

Same changes as Python 0.0.4, including the optional schema in the SDK, CLI and MCP server. Also:

  • KuzuStore options (max_db_size, buffer_pool_size, read_only, compression).
  • MCP tool arguments are validated, and close() lets an embedding process exit.

Docs: https://github.com/hydra-db/selfgraph/tree/main/typescript#readme

ts-v0.0.3

Choose a tag to compare

@arrry-hydraDB arrry-hydraDB released this 07 Nov 01:32

First public release of the SelfGraph TypeScript SDK. It reads and writes the same graph as the Python SDK.

npm install @hydradb/selfgraph

Same API, backends, CLI and MCP server as Python. Node 20+.

Docs: https://github.com/hydra-db/selfgraph/tree/main/typescript#readme