Initial Release

Better UI

bugfix

added afterinstall message

Initial Release

Initial Release

Initial Release

Initial Release

Initial Release

Initial Release

bump

Version bump

Version bump

Initial Release

processV9 — Release Notes

Changes over baseline (process.py)

Skip comment rows at read time

Uses skip_rows=365 and has_header=True instead of comment_prefix="#". Polars no longer scans and discards 365 header lines per file; it seeks directly to the data. The filter(source_id != "solution_id") guard is also eliminated since the comment block never enters the dataframe.

Largest-first file scheduling

Files are sorted by descending size before dispatch. On 8 cores (local CPU) with 20 files this follows LPT (Longest Processing Time) scheduling, minimizing the final idle-thread stall when the last batch of files completes.

Thread count capped at core count

max_workers = min(20, os.cpu_count()) — avoids 20 threads competing for 8 cores and reduces context-switching overhead compared to the hardcoded max_workers=len(files) in the baseline.

One polars thread per worker

POLARS_MAX_THREADS=1 prevents polars' internal Rayon thread pool from spawning additional threads per file read, which would over-subscribe the CPU when multiple Python worker threads are already running concurrently.

Initial Release

Initial Release

Made update for path issue in Docker

Changed the runtime Docker volume layout so that only data/out is bind-mounted. Nothing is precomputed, all work stays in the timed run.

Found some more optimizations. Using zgrep now, using csv.reader, some other small changes. Down to 378 meaningful code characters.

Added wrong about link

Initial Release

Added some optimizations for RunScript

Using docker mounts, moved file paths out of Python execution, bringing total RunScript.mac character count down to 452 characters, 433 of which are "meaningful code" (assuming "meaningful code" means "everything after ROUTINE RunScript")

Initial Release

Initial Release

add AI support

Initial Release

Release Notes: iris-vector-rag v0.11.4 (OEX 1.0.2)

This release represents a complete rewrite and production hardening of the
framework since the initial OEX listing (1.0.1).

Package renamed

Previously listed on OEX as intersystems-iris-rag (rag-templates on PyPI).
Now published as iris-vector-rag on PyPI with a matching OEX module name.
Importable as iris_vector_rag. The iris_rag/ shim preserves legacy imports.

Six pipeline strategies, unified interface

All pipelines share one factory call and one response shape — answer,
retrieved_documents, contexts, sources, metadata — swappable with a
single create_pipeline(type, ...) call:

• BasicRAG — dense vector retrieval
• BasicRAGReranking — retrieval + cross-encoder reranking
• CRAG — Corrective RAG with relevance-gated web fallback
• HybridGraphRAG — GraphRAG + vector hybrid with entity extraction
• MultiQueryRRF — multi-query generation fused via Reciprocal Rank Fusion
• ColBERT/PyLate (PLAID) — late-interaction ColBERT with in-DB PLAID index

ObjectScript SDK (RAG.SDK.*)

Five ObjectScript classes callable directly from IRIS without touching Python:

• RAG.SDK.Pipeline — run any pipeline strategy by name
• RAG.SDK.Search — BM25, IVFFlat, and vector search paths
• RAG.SDK.Schema — table initialization, schema status, pip install hook
• RAG.SDK.Bridge — overlay support and default table configuration
• RAG.SDK.Evaluate — RAGAS evaluation from ObjectScript

Connection layer overhauled

• All IRIS connections route through get_iris_connection() — one path,
  no duplicates.
• Auto-detects embedded runtime via iris.runtime.get().state — skips TCP
  entirely when running inside IRIS or when IRISINSTALLDIR is set.
• Lazy import iris.dbapi in connection_pool.py — no ImportError at
  import time when iris is absent.
• Unified via iris-embedded-python-wrapper — handles embedded-kernel,
  embedded-local, and native-remote backends transparently.

attach_existing_corpus

Zero-copy bridge: point any pipeline at tables already in IRIS without
re-ingesting data.

Validation layer

Pre-flight checks before any pipeline runs: required tables exist, embeddings
are ≥95% non-NULL, IRIS VECTOR format valid. auto_setup=True creates missing
tables and embeddings on first use.

RAGAS evaluation

Side-by-side pipeline comparison with faithfulness, context precision, and
context recall. Uses real PMC biomedical documents — no synthetic data.

REST API and MCP server

• FastAPI REST API ([api] extra) with Redis-backed sessions.
• MCP server ([mcp] extra) — all pipelines exposed as MCP tools, usable
  from Claude and other MCP clients.

iris-vector-graph integration

GraphRAG, ColBERT/PLAID, BM25, IVFFlat, and shortestPath delegate to
iris-vector-graph — IRIS-native graph and vector operations without leaving
the database.

IPM/ZPM install

zpm install iris-vector-rag now also runs
pip install iris-vector-rag==0.11.4 via RAG.SDK.Schema.Install().

CI/CD

• GitHub Actions CI on Python 3.11 and 3.12.
• Release workflow: tests → build → PyPI (OIDC trusted publishing) →
  GitHub Release with changelog notes attached.

First release on OEX!