Skip to content

News

Release notes and highlights. For a complete changelog, see the GitHub releases.

v0.10.5 — Load-Time Template Validation & Friendly Local-Model Errors

  • 🛡️ Load-time identifiers validation — a template whose node_id/relation_id/relation_members reference fields not declared in output now fails at Template.create with relation_id references '{type}' but relations.fields does not declare 'type', instead of crashing at first merge with an unrelated max() error. (production feedback; requires ontomem ≥ 0.6.1)
  • 🧠 Friendly local-model errors — without a config file, Template.create no longer silently falls back to OpenAI defaults; the error lists the three ways forward (he config init, -p vllm local setup, explicit clients).
  • ontomem 0.6.1 — exact-rollback safety: KeyExtractionError on 100% key-extraction failure, aggregated key_extraction_failed warnings, and remerge="mechanical" for deterministic (LLM-free, seconds-fast) rollbacks on large shared keys. (production feedback)

v0.10.4 — Visible Chunk Failures & Hypergraph Incremental Feed

  • 🐛 Hypergraph incremental feed fixed — feeding a second document into an existing hypergraph KA crashed with AttributeError: 'key_extractor' (_update_data_state referenced a nonexistent attribute). Now uses node_key_extractor/edge_key_extractor.
  • 👁️ Chunk extraction failures are visible — failed chunks were silently skipped with only a log line. Every feed_text()/parse() run now records them: read ka.extraction_failures (chunk index, stage, error). he feed/he parse print a warning when chunks were dropped.
  • ⚙️ on_error strategy — new constructor option: "skip" (default, current behavior) or "raise" to abort a feed on the first failing chunk. Available on all AutoTypes.

v0.10.3 — Working Cypher Export

  • 🐛 Cypher export fixed — he export cypher output could not be imported by Neo4j/Memgraph: statements lacked ; terminators (cypher-shell read the whole file as one query) and variables were re-declared within a single statement (Variable 'n' already declared from the second node on). Each node/edge/hyperedge is now its own ;-terminated statement, and hyperedge members use positional aliases (n0, n1, …). (#174)
  • 🧹 Consistency — AutoList/AutoModel raise the standard ValueError build-index hint instead of a bare Exception. (#175)

v0.10.2 — Scoped Chat, Method Guide & Persistence Fixes

  • 💬 Scoped chat — he talk --source/--tag (and ka.chat) narrow the context to attributed/tagged documents, with capability detection for non-tracked types. (#161, #160)
  • 📖 Choosing a method — new "baseline vs structured" guide plus a side-by-side chunk_rag vs graph_rag example on the same corpus, with cost/latency comparison. (#162, #163 — first two steps of #117)
  • 🏗️ Internal — GraphIndexMixin extracts the shared merge/index/search surface of graph & hypergraph (−168 lines). (#145)
  • 🐛 Fixes — temporal/spatial observation context now survives dump/load (was silently lost); MCP load failures return as tool results instead of crashing the stdio session; he config init no longer writes a broken same-provider embedder for LLM-only providers; Cog_RAG.chat forwards source_ids/tags; he list --lang zh keeps method templates; document templates get output-shape validation; CLI banner lists jsonld/cypher. (#167, #169, #165, #171, #159, #173, #155)
  • 📚 Docs — resolved leftover conflict markers in MCP docs; unsmashed JSON-LD/Cypher sections; type-aware index-layout notes; Gemini provider guide. (#149, #151, #153, #157)

v0.10.1 — Pickle-Free Index Storage

  • 🔒 Safe index format — AutoModel/AutoList indexes are now stored as JSON (index.json: vectors + documents) and rebuilt in memory on load — no pickle, no code execution. Legacy pickle indexes (<= v0.10.0) still load with a deserialization warning; he build-index --force migrates them. (#116)

v0.10.0 — More Export Formats, Gemini, and Quality Hardening

  • 📤 JSON-LD & Cypher exports — he export jsonld (pairwise edges + GraphML-style hyperedges) and he export cypher (Neo4j-compatible MERGE script; N-ary edges become Hyperedge nodes, not lossy cliques). (#139, #141)
  • 🔌 MCP export parity — export_graphml, export_csv, export_jsonld, export_cypher tools on he-mcp; MCP list_templates now includes methods and info shows chunks/timestamps/sources like the CLI. (#106, #131, #133, #144)
  • 📄 he feed accepts directories — parity with he parse: per-file source attribution, same skip/warning behavior. (#135)
  • 🧠 Native Gemini support — the previously unused [google] extra is now wired to create_llm: pip install "hyperextract[google]". (#137)
  • 🏗️ Internal — GraphIndexMixin extracts the shared merge/index/search surface of graph & hypergraph (−168 lines); contract tests for every registered method; test matrix now includes Windows. (#112, #108, #145)
  • 🛡️ Hardening — chunk options validated at parse time (chunk_overlap < chunk_size); relation_members requires source/target roles; Windows config.toml ACL restricted to the current user; FAISS deserialization warning. (#91, #92, #110, #94)
  • 🐛 Fixes — scoped search forwarded on hypergraph; he search prints nodes/edges instead of raw tuples; MCP search no longer crashes on 3-tuples; GraphML export respects --force. (#90, #144)

v0.9.1 — Scoped Search Completeness & Validator Hardening

  • 🐛 Fixes — scoped search forwarded on hypergraph (AutoHypergraph.search dropped source_ids/tags); AutoSet provenance wired end-to-end (recording, tags, rollback, persistence); Graph_RAG/Cog_RAG scope forwarding; friendly error for scope on non-tracked types.

v0.9.0 — Rich Document Ingestion & chunk_rag Baseline

  • 📄 Rich Document Ingestion — he parse / he feed now accept PDF, Word, PowerPoint, Excel, HTML, CSV/JSON/XML, EPUB and more via the optional ingest extra (pip install "hyperextract[ingest]", powered by MarkItDown). Non-UTF-8 text (GBK, etc.) is auto-detected; text-less (scanned) PDFs fail with a clear OCR hint instead of silently ingesting garbage.
  • 🧱 chunk_rag Baseline Method — a new zero-extraction method: documents are chunked and embedded as-is, and search returns raw text chunks. Zero LLM cost at ingestion, with full provenance (tags, scoped search, per-document rollback). The chunk-retrieval baseline for corpus Q&A and method benchmarking.
  • 🐛 Fixes — he tag crashed on every knowledge abstract (tag_source was never exposed on any type); he search/he talk/he feed/he remove --document crashed on method-built KAs (method/* templates were not resolvable from KA metadata).

v0.8.1 / v0.8.2

  • 🏷️ Source Tags & Scoped Search — he tag ./ka/ --source doc-1 --add legal, then he search ./ka/ "query" --tag legal to retrieve only within tagged documents. Works for graph, hypergraph, and set KAs. (#89, #84)
  • 🛡️ Input Validation — he parse / he feed now reject unsupported file types with a conversion hint instead of silently ingesting garbage. (#88)
  • 📦 Document Archive Fix — re-feeding the same source from a differently-named file no longer accumulates stale copies. (#89)

v0.8.0

  • 🔄 Document Upsert — re-feed an attributed document and its previous version is rolled back automatically: removed facts disappear, shared keys re-merge from surviving sources. (#84)
  • 📁 Per-File Source Attribution — he parse ./docs/ attributes each file by its name automatically; roll back or audit any single file later. An explicit --source still overrides.
  • ⏱️ Spatiotemporal Provenance — temporal/spatial/spatio-temporal graphs fully support source attribution and rollback, with deterministic (MERGE_FIELD) replay tests.

v0.5.0 – v0.7.0

  • 🗑️ Two-Tier Knowledge Deletion — hard-delete by key (he remove --node/--edge, orphan edges pruned) or remove a single wrong fact via LLM-assisted editing (he remove --edit-node --fact), with dry-run, key-invariance checks, and automatic backups. (#84)
  • 📜 Source Attribution & Provenance — he feed --source / he parse --source record each document's raw contributions; he remove --document rolls back exactly what one document contributed; he info --sources shows the ledger. (#84)
  • 📈 Incremental Everything — feed/parse/removal/edit patch the vector index in place (only affected vectors re-embedded); he feed skips documents whose content hash is unchanged (--refeed to force). (#84)
  • 🧪 he template validate — catch semantic template errors before paying for LLM calls: 9 diagnostic rules, --json for CI, --all for directories. (#77)
  • 📊 GraphML & CSV Export — desktop graph tools and spreadsheets; hypergraphs get a hyperedges table. (#85)
  • 🌐 OrcaRouter Provider — one key for 150+ models via create_client("orcarouter"). (#71)
  • 🔐 Config File Permissions — ~/.he/config.toml saved 0600. (#86)
  • 🔗 Obsidian wikilink fix — aliases no longer break on [ ] | # ^. (#87)
  • 🛡️ Chunk-Level Fault Isolation — one failed chunk no longer discards the rest of a multi-chunk extraction. (#78)
  • ⚡ MCP Python SDK 2.x — he-mcp works on mcp 1.x and 2.x. (#72, #82)
  • 🔀 Directed-Edge Fix — (source, target) order preserved; custom endpoint field names. (#74)
  • 🔑 DeepSeek API Key Fix — DEEPSEEK_API_KEY honored on the OpenAI-compatible path. (#76)
  • 🎓 Education Templates — course_concept_graph + curriculum_structure. (#80)
  • 🧭 Smaller Fixes — Graph_RAG.search 3-tuple; he talk -i --top-k; onboarding/docs overhaul. (#70, #73, #57)