DEV Community: RagLeap

I Built Tenant-Scoped Memory That Erases for Real — 89 Tests, 27 Mutations, Byte-Level Verified

RagLeap — Sat, 24 Oct 2026 16:00:33 +0000

ragleap-memory 0.1.0 + ragleap-security 0.1.0 — One SQLite File, Real Erasure, Scope Enforced in SQL

I just released two new libraries: ragleap-security 0.1.0 (101 tests) and ragleap-memory 0.1.0 (89 tests). Both v0.1.0, both first releases, both with mutation testing.

TL;DR: Tenant isolation enforced inside SQL, not in app code. And erase() actually removes bytes.

ragleap-security — stdlib only

from ragleap_security import Scope, AuditLog, SecretStr, check_size

scope = Scope.parse("org:gov-dept/region:south/branch:chennai")
print(scope.covers(Scope.parse("org:gov-dept/region:south"))) # True

child = scope.narrow("facility:factory1")

org:gov-dept/region:south/branch:chennai/facility:factory1

log = AuditLog("audit.jsonl")
log.append({"op": "add", "scope": str(scope)})
log.verify() # hash-chain check

secret = SecretStr("sk-...")
print(secret) # SecretStr(***) — never leaks in repr
check_size(data, max_bytes=1_000_000)

5 modules, no deps. You can audit it in an afternoon.

Why hash-chained JSONL? Tamper-evident. If someone edits line 10, verify() fails because hash chain breaks. Not tamper-proof (can't prevent edit), but can detect.

Verified: 101 tests, 30/30 mutations killed, reproducible build matches PyPI.

ragleap-memory — the interesting part

Problem with most memory libs: They do SELECT * FROM memories WHERE tenant_id = X in app code. But they still return different errors for "not found" vs "not allowed" — leaking existence.

Our approach: scope filtering inside SQL, same response for missing vs outside-scope.

from ragleap_memory import SQLiteStore
from ragleap_security import Scope

store = SQLiteStore("memory.db")
scope_a = Scope.parse("org:hospital/branch:chennai")
scope_b = Scope.parse("org:hospital/branch:mumbai")

store.add(text="ICU bed count changed to 12", type="fact", trust="trusted", scope=scope_a)

Same scope → found

print(store.search("ICU", scope=scope_a)) # [fact]

Different scope → looks like missing, not "access denied"

print(store.search("ICU", scope=scope_b)) # [] — same as missing

Typed entries: episode (raw events), fact (distilled), summary (compressed). All default to untrusted — you must explicitly trust.

Operations that matter for enterprise:

id1 = store.add(text="Old threshold", type="fact", scope=scope_a)
id2 = store.supersede(id1, text="New threshold is 80", scope=scope_a) # history kept

store.forget(id2, scope=scope_a) # soft delete
store.erase(id2, scope=scope_a) # hard delete + vacuum

After erase, byte-level check:

grep -a "New threshold" memory.db → not found

grep -a "New threshold" memory.db-journal → not found

We actually tested byte-level search after erase on two SQLite builds.

Other ops: export (dump for migration), sweep_expired (TTL), compact (vacuum), optional audit log without entry text (privacy).

FTS5 search: Full-text search, not vector. Why? Deterministic, no embedding cost, works offline. For Hindi, Tamil, Arabic, Chinese with spaces — whole-word tested. Without spaces (Chinese/Japanese) — not split yet, known limit.

Mutation testing

We run scripts/mutation_check.py — it mutates code (change == to !=, remove WHERE clause, etc.) and checks tests kill it.

security: 30/30 killed

memory: 27/27 killed

If a mutation survives, your tests are weak. We fixed until all killed.

Reproducible builds

Build locally before tagging, compute SHA-256, build on CI after tagging, compare. If different, fail. For these releases, local SHA matches PyPI SHA.

ragleap_memory-0.1.0-py3-none-any.whl sha256 c60a1e4f91e323c04ba04d7628abb7028039206d2330921a7ca00ae053a3533e
ragleap_security-0.1.0-py3-none-any.whl sha256 e3f31d3703d552cde53080e6d6c56cef733954d61deeaba586bdce48db53075d

Not verified — honest

No Windows / NFS / multi-process writes / power-loss tests
No retrieval benchmark (precision/recall unmeasured)
Tenant isolation holds inside lib, but nothing else uses it yet
No independent security review — see docs/threat-model.md

Install

pip install ragleap-security==0.1.0
pip install ragleap-memory==0.1.0

Repo: github.com/ragleap/ragleap-core

Release PRs: #907 (security), #922/#924 (memory)
Workflows: 40016023080, 40060689857

This is foundation for human digital twins and org-wide memory with real isolation. Next: integrate Scope into agents.

If you care about multi-tenant AI, try breaking the scope isolation — I'd love a failing test.

I Added Sub-Agents to ragleap-agents Without Breaking Anything — Here's How

RagLeap — Thu, 22 Oct 2026 06:30:54 +0000

ragleap-agents 0.5.0: Sub-agents — 208 tests, 8 live runs, 0 breaking changes

Tag: ragleap-agents-v0.5.0 (332b4c0) — 2 commits to main since.

TL;DR: make_delegate_tool lets a parent delegate to children, with taint flowing both ways and hard limits. Opt-in only.

The design goal

No breaking changes. Every addition is opt-in.

Old code still works

agent.run("do X")

New opt-in flags

agent.run("do X", tainted=True) # #828

Agent.run gained only *, tainted=False — keyword-only, defaults to old behavior.

make_delegate_tool — the new primitive

from ragleap_agents import make_delegate_tool, delegate_policy, Budget, MAX_DEPTH_CEILING

delegate = make_delegate_tool(
children={
"researcher": researcher_agent,
"analyst": analyst_agent
},
# limits per root run
)

parent = Agent(tools=[delegate, send_message, ...])
result = parent.run("Research competitor pricing and draft a note")

Implementation in #827, #843, #849:

Child runs in isolated context (its own tools, budget)
Returns as untrusted, fenced observation — parent must treat as untrusted
Taint: if parent.tainted → child.tainted=True, if child.tainted → parent.tainted=True
Budget: Budget + MAX_DEPTH_CEILING cap depth and cost per root run

Why fenced? To prevent prompt injection via child answer becoming parent's instructions.

Child approval model (v0.5.0 limit)

Child proposes send_message → needs approval

In 0.5.0: does NOT pause parent, stops with approval_not_available

This is intentional for safety. No privilege escalation from child to parent approval queue.

Future: explicit child approval queue with owner confirmation on host (Phase 2 permission system).

current_run()

from ragleap_agents.context import current_run

def my_tool(x: str):
ctx = current_run()
print(ctx.run_id, ctx.taint)
return x

Useful for custom logging / policy.

Verified (with numbers)

208 tests pass on Python 3.10 (VPS c31ab4e), CI green on 3.10/3.11/3.12
Wheel: exposes make_delegate_tool, delegate_policy, Budget, MAX_DEPTH_CEILING, Agent.run(..., tainted=False)
Live: scripts/live_delegate.py #890
- Provider: Groq openai/gpt-oss-20b, native mode
- 8 runs (4 tasks × 2 configs)
- Delegation: 8/8
- Tainted: 8/8
- send_message blocked: 8/8 (approvals never granted)
- Taint gate held: 3 runs proposed send after child read note → held for approval

Not verified — honest

Injection resistance: 2/8 read injected note unprompted → no send. Not enough for claim.
Child stop path tested only with scripted models in live runs.
2/8 runs failed on unparseable_reply / invalid_plan — same format intermittency as 0.4.x
No benchmarks, no 3.11/3.12 runtime check on published wheel, no resume test for 0.4.x saves.

We ship honest limits, not marketing.

Install & Provenance

pip install ragleap-agents==0.5.0

Digests:

whl: 8961d9ff4c131313b899dd8907f68fa2fa7672dd642a5e0ff34cb6c992334712
tar.gz: 12a1204bbe0c88e9afb45c89ab5c11f59dde33b2851dedb022c032bdefaa14c0

Release PR #892, workflow 39653944931, 2026-10-21 18:35 UTC. Trusted publishing with Sigstore.

What's next

This unlocks Phase 3 of RagLeap roadmap:

Lead agent that splits task → parallel children → citation merge
Budget per child + cancellation
Multi-provider live tests

Code: github.com/ragleap/ragleap-core

If you build multi-agent, try the delegation API and tell me where it breaks.

ragleap-core v0.11.0 – Documents Tab, One Scheduler, and Early Computer Control

RagLeap — Tue, 20 Oct 2026 16:20:28 +0000

ragleap-core v0.11.0 is out – first release under `ragleap` org

Repo moved: antonyrag/ragleap-core → ragleap/ragleap-core
Tag: v0.11.0 / commit c1fc164

This is a stability + product release. Here's the breakdown.

1. Documents Tab (the main feature)

Before: adding docs was scattered.
Now: dedicated Documents tab to list, upload, and delete.

Important: deleting also removes the document's entries from the knowledge graph.


python
# #710, #711 – deletion now cleans graph
# Dashboard now accepts:
# 28 file types + web pages by address (#720)

2. AI Office

Two new tabs:

Proactive
Connectors

PR #740 – foundation for AI employees to act, not just chat.

3. Offline Language Detection

LANGUAGE_DETECTION_MODEL=lite

Works without internet now (#741). Useful for air-gapped installs.

4. Infra fixes that matter

Non-root container:

# #742
USER 10001
# pinned dependencies

One scheduler:
Previously, extra web workers repeated scheduled triggers. Now only one process runs background jobs (#817).

5. Providers

DemonRoute added – 20 providers total (#851).

# provider count
20

6. Experimental: Computer Control (OFF by default)

⚠️ Not independently reviewed. Do NOT enable for real data.
Design:

Share via CLI only: ragleap files
Sandbox runner = read-only
AI can: list, read, ask for access
Access request ≠ grant. Always needs owner approval.
File contents never enter activity log
Grants visible in Files tab
ragleap files off = pause everything

Issues: #747, #776, #777, #789-#792. Review RFC: #800.

Community

Testing issues #779-#787

Number words for every language #801

ragleap-phone RFC #814
Thanks @sammist2 @mig-builds @ahmet-izgi @haterfactory

Upgrade

Container user is now 10001. No migration needed, app writes no files of its own.
Packages (ragleap-rag, ragleap-graph, voice, agents) versioned separately.

Known Limits

Neo4j default password still public in compose (localhost only) – per-install random password planned

Tested: Windows laptop + Linux servers. NOT verified: macOS, Linux desktops, Ollama on Windows (#779-#781)

No named users/roles, no security audit yet

Links:

Release: github.com/ragleap/ragleap-core/releases/tag/v0.11.0
Core: github.com/ragleap/ragleap-core

If you self-host, try the Documents tab and tell me what breaks. PRs welcome.

Testing Python-to-Java Parity in a Graph Retrieval Library

RagLeap — Fri, 16 Oct 2026 05:00:15 +0000

What does it take to port a graph-retrieval library from Python to Java without relying on API similarity alone?

For RagLeap Graph Java v0.2.0, the answer was a combination of parity tests, deliberate mutations, database integration tests, and validation of the actual published artifact.

Release: RagLeap Graph Java v0.2.0

Features in this release

The Java library now includes:

Optional LLM-based entity and relation extraction.
Configurable extraction and entity deduplication.
Cross-chunk relation handling.
Hybrid vector + graph retrieval and graph-only retrieval.
Chunk and graph-path citations.
Pluggable graph and vector search interfaces.
Backfill utilities and a schema migration runner.

Regex-based extraction remains the default. LLM integration is optional.

How we approached parity testing

We compared the Java implementation against the Python ragleap-graph v0.9.0 implementation.

Rather than comparing method names, we exercised behavior using recorded calls, parameters, outputs, and scripted model responses.

The test scenarios covered:

422 string-similarity pairs.
25 extraction-configuration cases.
Entity and relation extraction.
Graph upserts, including Cypher statements and parameters.
Backfills and migration behavior.
Retrieval scenarios and retrieval configurations.

We also ran 27 deliberate mutations across the new functionality. Every injected break was caught by the parity tests.

This is evidence for the tested cases, not a mathematical guarantee of complete equivalence.

Why mutation testing matters

A passing test suite can still miss bugs if its assertions are too weak.

Mutation testing deliberately changes behavior to check whether the tests notice. For this release, the 27 deliberate breaks were caught by the parity tests.

That gave us additional confidence that the tests were checking behavior rather than merely executing code paths.

Testing the real release

The full local run used generated parity fixtures and a live Neo4j 5.26.29 Community instance:

129 tests
0 failures
0 skipped tests

After publishing, we resolved the Maven artifact in a clean project and exercised health checks, upserts, backfills, and graph-only retrieval against live Neo4j.

Maven coordinates:

io.github.antonyrag:ragleap-graph:0.2.0\

What still needs testing?

LLM extraction hasn't been verified against a real model provider. Compatibility with other Neo4j deployments also needs more testing, and the optional vector-search adapter hasn't been tested against a live RagLeap facade.

The PostgreSQL audit writer hasn't been ported yet.

Documenting these limitations is part of making a library easier to evaluate responsibly.

Try it and share feedback

Source and release notes:

https://github.com/antonyrag/ragleap-core/releases/tag/ragleap-graph-java-v0.2.0

RagLeap is fully open-source and self-hosted, with a wider ecosystem of retrieval, graph, and AI tooling libraries.

If you work with Java or Neo4j, I'd appreciate feedback on the API, compatibility, test coverage, and integration experience.

What integration test would you add before using this library in production?

ragleap-rag Java v0.13.0 and ragleap-graph v0.1.0 are on Maven Central — Full Python Parity, 502 Tests

RagLeap — Thu, 15 Oct 2026 08:16:31 +0000

Java devs — RagLeap is now native Java, on Maven Central!

No Python service needed anymore.

Quick Start

pom.xml:

io.github.antonyrag
ragleap-rag
0.13.0

io.github.antonyrag
ragleap-graph
0.1.0

Code:


java
var rag = new RagLeap(config);

// Ingest
rag.ingest(Path.of("report.pdf"));
rag.ingestText("RagLeap is safety-first RAG");
rag.ingestImage(Path.of("chart.png"));

// Ask
String ans = rag.ask("Summarize report with sources");

// Stream
rag.askStream("Explain", System.out::print);

// History
rag.getHistory();
rag.clearSession();

// Documents
rag.listDocuments();
rag.deleteDocument("doc-id");


Graph:

var graphConfig = new GraphConfig("http://localhost:7474", "neo4j", "password");
var graph = new GraphIndex(graphConfig, cypherRunner);

graph.upsertDocument(docId, text);
var docs = graph.findDocumentsByEntities(List.of("RagLeap", "Antony"));
var lineage = graph.findLineage("entity-id");



How we proved parity

Most ports test final output. We tested ordered behavior.

* 33 scenarios vs Python RagLeap 0.13.0 (d210dab)
* Fake embedder, vector store, generator, memory, hooks
* Full ordered log must match — calls, hooks, memory writes
* Mutation test — change behavior → test fails

**Results:** 502 tests — 0 failures.

**Graph:** 252 logic + 14 upsert + 47 query vs Python 0.9.0 (226fa4d). Cypher + params (int vs float!) must match.

**Live:** Tested against Postgres and Neo4j 5.26.29 Community by resolving jar from Maven Central in clean project.

**

## What's NOT ported (honest)
**

**ragleap-rag:** reranking, Redis cache, ingest_url, async variants. QueryRewrite is enum, streaming is callback. ingestBatch keeps metadata (Python drops it for file items).

**ragleap-graph:** LLM extraction, deduplication, GraphRetriever, backfills, migration, Postgres audit writer.

Check release notes for full list — no surprises.


**

## Roadmap
**


* reranking + Redis cache
* GraphRetriever (hybrid vector+graph)
* LLM entity extraction in Java


Contributions welcome — raise issue or PR on ragleap-core.

Try it: mvn dependency:get and let me know what breaks!

RagLeap Core v0.10.1 — I shipped 2 releases in 44 mins after Windows broke everything

RagLeap — Tue, 13 Oct 2026 07:08:47 +0000

I shipped ragleap-core v0.10.0 with login throttling, Setup tab, and a read-only skill for coding agents. 44 minutes later I shipped v0.10.1.

Why? First real Windows test with Docker Desktop.

A local PostgreSQL was already on port 5433. Installer stopped. That's it.

That's the most honest bug report you'll get in AI.

What shipped in v0.10.0 (hardening):

Login throttling: 10 bad API keys in 5 mins → 15 min 429 lock with Retry-After. Proxy-aware via TRUSTED_PROXIES, off with AUTH_THROTTLE=off
Setup tab + GET /setup/status: Shows what's configured/missing, never shows secrets, zero AI calls
Read-only skill for coding agents: manifest + small client (health, employees, ask). Can't change anything by design
Ollama context 8192 default: Long docs no longer silently cut
Installer checks: Free disk check (4GB / 14GB for local AI), refuses if Compose project name already used, can use Ollama app already running on Win/Mac. Added install.ps1 PowerShell installer
Security: Every third-party GitHub Action pinned to exact commit

What shipped in v0.10.1 (Windows reality):

This is where it gets real.

Free host ports: All ports are now settings — APP_PORT 8000, DB_PORT 5433, VOICE_PORT 8766, NEO4J 7475/7688, REDIS 6380. On fresh install, installer picks next free port if busy, writes to .env, uses it for dashboard link. Existing .env never changed.
ragleap key --rotate: Writes new random API key to .env, restarts only app, checks old key is refused + new accepted, prints new link. DB untouched.
Browser instead of printed key: On desktop (Win/Mac/Linux with display, no SSH), opens dashboard in browser, no key in terminal. Servers/SSH/CI still print link. RAGLEAP_NO_BROWSER=1 to always print.
Fixed: DB exists check matched ANY Docker volume with same name, even unrelated installs → now matches only this setup's volume. API docs now says needs x-api-key header (401 by design).

Honest limits (we list these in release notes):

Tested on one Windows machine with Docker Desktop, Gemini/Ollama/skip menu set to skip. Ollama on Windows not tested yet
One-liner irm ... | iex fails on that machine, cause unknown. Use two-step: $s = irm ; iex $s
PATH hint writes for bash, not PowerShell
Windows reserved excluded ranges not detected by busy-port check
While browser launches, key briefly visible in process list

What is RagLeap Core?

Self-hosted AI Business OS — 46 autonomous AI Employees (9 core + 37 vertical like Recruiter, Real Estate Agent, Legal Intake), Graph RAG (Neo4j), multi-channel (WhatsApp/Telegram/Discord/Voice), BYOK 19 providers + Ollama local, MIT licensed.

curl -fsSL https://ragleap.com/install.sh | bash

or Windows

$s = irm https://ragleap.com/install.ps1; iex $s

18 stars → aiming 20⭐️

GitHub: github.com/antonyrag/ragleap-core
Website: ragleap.com

Building in public

ragleap-agents v0.1.0: A Fail-Closed Act-Observe Loop in 200 Lines

RagLeap — Sun, 11 Oct 2026 08:03:59 +0000

I just shipped ragleap-agents v0.1.0. It's not a framework, it's a guardrail.

llm is any callable(prompt) -> str. No SDK lock-in.

The loop

llm -> JSON action -> validate against JSON Schema -> check ToolPolicy -> execute or pause for approval

Security model


python
ToolPolicy(
  taints=True,      # this tool reads untrusted data
  outbound=True,    # this tool sends data out
  requires_approval=True
)

Rule: After a tainting tool has run, every later outbound tool needs approval.

No policy? Counted as both taints + outbound. Fail closed.

Tool outputs are capped, fenced, tags neutralized case-insensitively.


**

## Pause / Resume
**

StateStore is pluggable. InMemoryStateStore included.

Approve -> sends. Reject -> never sends. Replay -> ResumeError.

**

## Provenance you can check
**

PR #628 merged as d3e0c14, release PR #646 -> tag 0482515
CI 37889921507 success, publish run 37890106262 with trusted publishing.

Wheel: d305238f... Tar: 44cf1267... - identical in logs and PyPI.

41 tests, mutation-checked.

**

## Install
**

pip install ragleap-agents==0.1.0
pip install ragleap-tools==0.4.0  # dependency

This does NOT stop prompt injection. It limits blast radius IF your policies are accurate.

Code: PR #628 in antonyrag/ragleap-core

RagLeap v0.9.0 — Self-hosted AI Office with sandboxed agents, approval inbox, and pluggable embeddings

RagLeap — Thu, 08 Oct 2026 12:56:27 +0000

RagLeap is an open-source AI-employee platform. v0.9.0 ships the AI Office.

TL;DR

Dashboard at /office
ragleap CLI: launch, stop, status, logs, update (with DB backup), key
Pluggable embeddings (Ollama, OpenAI, Mistral, Gemini) + vector size check
Agent loop: act-observe, sandboxed exec, MCP client, cron triggers
Approval system: off/semi/full, taint after web fetch, network-less sandbox

How settings work now

Settings page > .env. UI shows source of each value. Keys are encrypted at rest, write-only in UI. Test connection button included.

Safety by default

API-key auth on all data routes, no stack traces to client
SSRF-safe fetch, sandbox has no network
Approval inbox for all tool calls in semi/full mode

Install

curl -fsSL ... | bash

or

pip install ragleap-core
ragleap launch

Open PRs and issues welcome. I need clean-machine testers for installer.

Repo: github.com/antonyrag/ragleap-core
Release: v0.9.0 - f11d458

We Shipped ragleap-terraform v0.1.0: How a Solo Dev is Building DevOps for ANY Software — Live-Verified, Not Hype

RagLeap — Tue, 06 Oct 2026 06:42:49 +0000

We Shipped ragleap-terraform v0.1.0: How a Solo Dev is Building DevOps for ANY Software

I shipped ragleap-terraform v0.1.0 8 hours ago. It's our 9th library. And it changes what RagLeap is.

RagLeap started as RAG. Now it's DevOps for ANY software. Here's the honest breakdown — what's live-verified, what's next, and what we won't claim.

TL;DR

New: ragleap-terraform 0.1.0 — kind cluster + full observability stack via Terraform, verified on a real Docker host
Done: 4 infra libs — ops, app-chart, observability, terraform — all live-verified
Philosophy: "enterprise-grade foundations, live-verified" — NOT "proven at global scale". One node, single-fault tests. We ship truth.

🚀 The Latest: ragleap-terraform v0.1.0

What it does:
A Terraform module that creates a local kind cluster and installs ragleap-observability with helm_release.


hcl
module "kind-cluster" {
  source = "your-registry/ragleap-terraform"
  cluster_name = "ragleap-dev"
  grafana_admin_password = var.grafana_password
}


**Variables:** cluster_name, namespace, install_observability, observability_chart_path, grafana_admin_password, install_postgres_exporter (default false), helm_timeout, kubeconfig_path

**What's inside:** examples/kind ships inside the wheel. pip install ragleap-terraform and you have it.


**Verified — Not Mocked:**

Applied and destroyed on a real Docker Desktop host with Terraform 1.9.8. All chart pods Ready. Your kubectl context untouched. This is why I say "live-verified".

**Known Limitations (honest):**

Local kind only. No AWS/GCP/Azure modules yet — those need a real cloud account to verify, not just code.

postgres-exporter off by default — needs ragleap-ops and a Secret no chart creates yet.

State and kubeconfig hold credentials in plain text. Do not interrupt apply/destroy.

This is v0.1.0 for a reason. It works, it's verified, and we're transparent about gaps.

**

## ✅ What's Done & Live-Verified So Far
**

**1. ragleap-ops 0.4.1**

Helm chart, SecurityContext, read-only root filesystem (not neo4j), zero-permission ServiceAccounts, NetworkPolicies, Ingress + TLS, multi-env values, Postgres and neo4j backup/restore with RTO, runbooks drilled against broken clusters.

This wasn't helm install and hope. We broke clusters on purpose and drilled restore.

**2. ragleap-app-chart 0.2.1**

Generic Helm chart for any developer's services. If you have a service, this deploys it with the same security standards as ragleap-ops. This is the "ANY software" part.

**3. ragleap-observability 0.6.0**

Prometheus v3, Grafana 12, Loki 3.6, Promtail, AlertManager, Slack and email receivers (delivery proven against a stand-in), pods-only RBAC, configurable Grafana password, Trivy scans of all images.

We started at 383 findings. Down to 14 on four images. postgres-exporter still at 48, Promtail at 86 — next bump cycle will fix.

**4. Release Process**

PyPI keywords, TestPyPI → PyPI, tags, GitHub Releases, docs surfaces. Boring but critical.

**

## 🔜 What's Next — Ordered by Level
**

I organize everything by solo → team → corporate. You only need what you need.

#**## SOLO DEVELOPER (Next 2 weeks):**

* Finish Loki retention → ship 0.6.1
* Real Slack/email send — right now only stand-in is proven
* Automate postgres-exporter Secret + DB role — no chart does it today, it's manual
* ragleap-ansible for VPS bootstrap (never started)
* Image bumps

**

## TEAM (Next month — needs your sign-off):
**

* HPA and PodDisruptionBudgets in ragleap-ops + app-chart
* Helm chart testing in CI (ct) + scheduled Trivy job — both edit CI files
* GitOps (ArgoCD or Flux) — this is the blocker for canary and blue-green
* SLO and SLI dashboards
* ServiceAccounts, RBAC and PDBs for app-chart


**

## CORPORATE (Later — needs cloud accounts & audits):
**

* Cloud Terraform modules (AWS, GCP, Azure) — they'd be unverified without an account, so I won't fake it
* Multi-AZ, multi-node, cluster autoscaling (blocked — kind is one node)
* Ingress rate limiting + WAF, audit logging, image SBOM + provenance
* Tracing, SSO for Grafana, chaos testing
* Compliance (SOC 2) — needs external audit, not just engineering



## Suggested Order:

Loki retention → 0.6.1 → HPA/PDB → Ansible → Trivy/ct in CI → GitOps

**

## 🎯 The Honest Claim
**

This is important. Even with every item done, the accurate claim is:

**

## "enterprise-grade foundations, live-verified"
**

It is NOT "proven at global scale". Everything has run on one node, with single-fault tests.

I'm a solo dev building in public from Chennai. I won't claim what I haven't tested. Many startups claim "enterprise-ready" after one Helm chart. I drilled backup/restore against broken clusters before claiming RTO.

**

## 📦 Where We Are
**

46 AI Employees — modular RAG & self-thinking intelligent agents
9 Libraries — 8 Python + 1 Java
45k+ Downloads
MIT Licensed, Self-Hosted, BYOK — 17 LLMs + Ollama local
45+ Vector DBs


Links:

Core: github.com/antonyrag/ragleap-core (17 stars, help us reach 50?)
Docs: docs.ragleap.com
Packages: packages.ragleap.com
PyPI: pypi.org/project/ragleap-terraform

If you want DevOps that doesn't hide behind marketing, try:

pip install ragleap-terraform

Feedback welcome. Issues welcome. Stars appreciated.

Building in public. Shipping truth.

— Antony

I benchmarked 4 vector DBs with identical vectors — retrieval was identical, scores were not

RagLeap — Mon, 05 Oct 2026 03:40:24 +0000

A comment on my last release: backend portability only matters if retrieval quality stays comparable. Fair.

So I built ragleap-rag/benchmark.

TL;DR:

296 SQuAD paragraphs, 60 questions, same vectors in pgvector, FAISS, Qdrant, Weaviate vs exact float64 scan
overlap@10 = 1.000 on all four
Found 5 bugs: score scale mismatch + 3 edge cases
Fixed in v0.11.0

How I ran it:

cd java/ragleap-rag
./benchmark/run.sh # starts pgvector (docker), FAISS in-mem, Qdrant, Weaviate

generates benchmark/RESULTS.md + results.csv

Core logic:

var groundTruth = exactScan.query(queryVector, 10);
for (var backend : List.of(pgvector, faiss, qdrant, weaviate)) {
var results = backend.query(queryVector, 10);
double overlap = overlapAtK(groundTruth, results, 10);
double scoreDev = meanAbs(results.scores - (cosine+1)/2);
}

The bug that matters:

FAISS IndexFlatIP returns inner product = cosine if normalized. Weaviate returns cosine distance. Both in [-1,1].

But my abstraction promised [0,1]. Threshold-based filtering broke silently.

Fix:

// In FaissStore & WeaviateStore
return (rawScore + 1.0) * 0.5;

Now deviation: 0.17561 -> 0.00003

Other 3 bugs:

insert([0.1, 0.2]) into dim=768 store — should throw IllegalArgumentException
search() on empty Weaviate collection — should return [], not 500
Failed batch write left ragleap_docs row without vectors

Added 19-case matrix in FailureModeTest.java.

What's next?

Milvus/Pinecone via Testcontainers if I get CI keys
100K Wikipedia scale test for recall@10 vs efSearch

Repo: https://github.com/antonyrag/ragleap-core
Results: java/ragleap-rag/benchmark/RESULTS.md
Maven: io.github.antonyrag:ragleap-rag:0.11.0

If you use vector DB abstraction, test your score normalization. Overlap is easy, scores lie.

RagLeap Triple Release: Fixing Fresh Install Breakage, Adding Slack Alerts & GitHub Search

RagLeap — Fri, 02 Oct 2026 12:54:32 +0000

Shipped 3 packages yesterday/today. Here's the technical breakdown.

ragleap-rag v0.12.10

Stack trace:

ModuleNotFoundError: No module named 'lxml.html.clean'

Root cause: lxml moved clean to lxml_html_clean==0.4.0.

Fix in pyproject.toml:

[project.optional-dependencies]
web = ["trafilatura>=2.2.0", "lxml_html_clean>=0.4.0"]
formats = ["pandas>=2.2.2", "pyarrow>=14.0.0", "xlwt>=1.3.0"]
all = ["ragleap-rag[web,formats,qdrant,weaviate,pinecone,milvus]"]

Problem 2: _extract_parquet calls Table.to_pandas() — needs pandas, but wasn't in extras.

Now raises proper ValueError with install hint if missing.

CI Fix: Before → pip install.[test] only → 63 tests skipped.
Now → pip install.[formats][web][qdrant][weaviate][pinecone][milvus][test] → 308 passed.

Guard test:

def test_all_extras_installed_in_ci():
if os.getenv("CI") == "true":
import qdrant_client, weaviate, pinecone, milvus

ragleap-observability v0.4.0

Helm values:

alertmanager:
receivers:
slack:
enabled: true
api_url_file: /etc/alertmanager/secrets/slack_api_url
email:
enabled: true
smtp_auth_password_file: /etc/alertmanager/secrets/email_password

Secret mount is read-only. Tested with prom/alertmanager:v0.27.0 image + amtool check-config.

ragleap-tools v0.3.0

Zero-dep GitHub search:

class GitHubSearchConfig(BaseModel):
token: Optional[str] = None
num_results: int = 10 # clamped 1-20

def make_github_search_tool(config: GitHubSearchConfig):
# uses urllib.request, sets User-Agent: ragleap-tools/0.3.0
# handles 422 gracefully -> success=False

Live test without token:

GET https://api.github.com/search/repositories?q=rag
Rate limit: search=10, core=60 (docstring fix pending)

Install all 3:

pip install ragleap-rag==0.12.10 ragleap-observability==0.4.0 ragleap-tools==0.3.0

Repo: github.com/antonyrag/ragleap-core — star if useful ⭐

v0.7.5 — Maintenance Snapshot: 5 Packages, 477 Tests Green

RagLeap — Tue, 29 Sep 2026 06:45:56 +0000

No core changes in this one — but if you maintain a monorepo, you know this is the important release.

v0.7.5 is a snapshot since v0.7.0

If you use only ragleap-core, stay on v0.7.0. This is for monorepo users.

What's bundled:

1. ragleap-observability v0.3.0

AlertManager + first real alert rule — verified with curl /api/v1/rules = ok
Prometheus fix: strategy: Recreate to avoid PVC crash-loop
Loki Deployment + Grafana datasource wiring

2. ragleap-tools v0.2.0


python
from ragleap_tools import search_web
# BYOK - plug your own provider
result = search_web("ragleap observability", provider="tavily")




**

## 3. ragleap-rag
**

ingest(metadata={"source": "docs"}) — now supported
Fix: QdrantBackend cosine score normalization
Fix: WeaviateBackend custom port + named-vector search

## **4. Java Port v0.10.0**

+4 backends: Qdrant, Weaviate (live-verified), Milvus, Pinecone

## **5. Housekeeping**

clone-count badge automation (real traction > stars)
SBOM regenerated for compliance
477 tests passing

Why we care about SBOM + badges? Enterprise adoption needs it.

Repo: github.com/antonyrag/ragleap-core

DEV Community: RagLeap

I Built Tenant-Scoped Memory That Erases for Real — 89 Tests, 27 Mutations, Byte-Level Verified

org:gov-dept/region:south/branch:chennai/facility:factory1

Same scope → found

Different scope → looks like missing, not "access denied"

After erase, byte-level check:

grep -a "New threshold" memory.db → not found

grep -a "New threshold" memory.db-journal → not found

I Added Sub-Agents to ragleap-agents Without Breaking Anything — Here's How

Old code still works

New opt-in flags

Child proposes send_message → needs approval

In 0.5.0: does NOT pause parent, stops with approval_not_available

ragleap-core v0.11.0 – Documents Tab, One Scheduler, and Early Computer Control

ragleap-core v0.11.0 is out – first release under ragleap org

1. Documents Tab (the main feature)

Testing Python-to-Java Parity in a Graph Retrieval Library

Features in this release

How we approached parity testing

Why mutation testing matters

Testing the real release

What still needs testing?

Try it and share feedback

ragleap-rag Java v0.13.0 and ragleap-graph v0.1.0 are on Maven Central — Full Python Parity, 502 Tests

Links

Quick Start

RagLeap Core v0.10.1 — I shipped 2 releases in 44 mins after Windows broke everything

What shipped in v0.10.0 (hardening):

What shipped in v0.10.1 (Windows reality):

What is RagLeap Core?

or Windows

ragleap-agents v0.1.0: A Fail-Closed Act-Observe Loop in 200 Lines

The loop

Security model

RagLeap v0.9.0 — Self-hosted AI Office with sandboxed agents, approval inbox, and pluggable embeddings

TL;DR

How settings work now

Safety by default

Install

or

We Shipped ragleap-terraform v0.1.0: How a Solo Dev is Building DevOps for ANY Software — Live-Verified, Not Hype

We Shipped ragleap-terraform v0.1.0: How a Solo Dev is Building DevOps for ANY Software

TL;DR

🚀 The Latest: ragleap-terraform v0.1.0

I benchmarked 4 vector DBs with identical vectors — retrieval was identical, scores were not

generates benchmark/RESULTS.md + results.csv

RagLeap Triple Release: Fixing Fresh Install Breakage, Adding Slack Alerts & GitHub Search

ragleap-rag v0.12.10

ragleap-observability v0.4.0

ragleap-tools v0.3.0

v0.7.5 — Maintenance Snapshot: 5 Packages, 477 Tests Green

What's bundled:

ragleap-core v0.11.0 is out – first release under `ragleap` org