Updates from RagLeap
Announcements, honest write-ups, every release with its notes, and the project's guides — all sourced from the real repository.
Featured
v0.13.6: a sub-agent that needs no tool answers its task, and a slow provider no longer stalls planning
ragleap-core v0.13.6 fixes two bugs found by the first real-model test of sub-agents (off by default): a child with no tool now answers, and one provider call has a 45-second limit.
Outside security assessment of RagLeap Core: Aleksei Musienko (AMQIC) and what we fixed
Who assessed RagLeap Core from the outside, what he did, which findings we confirmed and fixed in v0.13.4 and v0.13.5, our wider security work, and what is not done. Not a certification.
v0.13.5: old approvals expire, Discord requests with an old timestamp are refused, and connector credentials fail closed
ragleap-core v0.13.5 expires approvals older than seven days by default, refuses Discord requests with an old timestamp, and refuses to store connector credentials in plain text without a key.
v0.13.4: security fixes after an outside assessment (approvals run once, file actions always need approval, chat approvals only from a private chat)
ragleap-core v0.13.4 fixes four findings from an outside static assessment by Aleksei Musienko (AMQIC): approvals run once, file actions always need approval, chat approvals only from a private chat, and credentials out of logs.
Encrypted export and import for LLM memory: AES-256-GCM, keys and what it does not protect (ragleap-memory 0.3.1)
How ragleap-memory 0.3.1 encrypts memory export files with AES-256-GCM through ragleap-security, with a key, a key ring or a passphrase, and what the encryption does not protect.
Audit logs for AI agents: what to log, what to keep and what not to claim (ragleap-core v0.13.1 to v0.13.3)
What an AI agent audit log can contain, how RagLeap Core v0.13.1 to v0.13.3 logs actions as JSON lines with hashes, filters and opt-in retention, and what it does not cover.
Prompt injection in AI agents: what a tainted run is, and what a 100-run test with two real models showed
How RagLeap Core v0.13.0 treats ingested documents as untrusted for actions, what a tainted run is, and what a small 100-run test did and did not show.
How to build an AI agent that asks for approval before it acts (Python, ragleap-agents 0.5.1)
Add human approval to a Python AI agent with ragleap-agents 0.5.1: pause before outbound tools, resume on a human decision, and what it cannot stop.
How to isolate tenants in a RAG app: for_tenant, vector-store filters and the Redis gap (ragleap-rag 0.16.0)
How to keep each customer's documents apart in a Python RAG app: tenant views in ragleap-rag 0.16.0, filters in ragleap-vectorstores 0.7.0, run scope in ragleap-agents 0.5.1, and what is not covered.
v0.13.3: agents can propose writing and deleting files with your approval (off by default), a signed core image and dependency fixes
ragleap-core v0.13.3 lets agents propose writing and deleting files, always with approval and behind switches that are off by default. Also a signed core image and dependency fixes. What is not covered yet.
RagLeap Core: self-hosted AI employees that ask before they act. What it is, what it is not and what is verified
RagLeap Core is an open-source (MIT), self-hosted engine for document-grounded AI employees. What it does today, how to install it, what it is not (single-tenant today, multi-tenant planned), what others have tested and what is not verified.
v0.13.2: audit filters, opt-in audit retention, request ids in logs and a sign-in rate limit (audit and sign-in hardening)
ragleap-core v0.13.2 hardens audit and sign-in: filterable audit reads, opt-in retention, request ids in log lines and a rate limit on single sign-on. What is on by default, what is not, and what is not covered yet.
ragleap-graph 0.10.0: deterministic community detection for a RAG knowledge graph on Neo4j
ragleap-graph 0.10.0 adds deterministic Louvain community detection for the entity graph. The README quickstart, how communities are stored and retrieved, and the known limits.
ragleap-rag 0.15.0: evaluate retrieval with ranking metrics, and what changed for Python 3.11
ragleap-rag 0.15.0 adds ranking metrics (hit@k, recall@k, precision@k, nDCG@k) to its evaluation and requires Python 3.11. The README quickstart and evaluation examples, what changed and what is not verified.
ragleap-voice: find, lock and verify open speech models safely (pre-alpha, every model unverified)
ragleap-voice is a pre-alpha library for local-first speech tooling. How it finds, locks, downloads and verifies speech models, what it measured on one clip, and what is not verified.
ragleap-memory 0.3.0: vector and hybrid search for LLM memory in one SQLite file, with your own embedder
ragleap-memory 0.3.0 adds optional vector and hybrid search over vectors your own embedder produces. How it works, what leaves the process, how to upgrade safely, how it was tested and what is not verified.
v0.13.1: structured audit log lines for every agent action, plus a security incident runbook (a small release)
ragleap-core v0.13.1 is a small release: every agent action now also writes one structured JSON audit line. What the lines contain, how to turn them off, what is not covered yet, and what else is in the release.
ragleap-optimizer: measure token savings on your own cases before you trust them (alpha, not independently audited)
ragleap-optimizer 0.1.0 reduces the token use of LLM requests and reports what each step changed and whether it can lose information. How to use it, the one lossy step, the cache, how to evaluate it and its limits.
ragleap-trace: dependency-free tracing for a Python RAG app, with privacy built in and the limits stated
ragleap-trace 0.1.0 records spans in your own process with hard limits and no content by default, and has an adapter for ragleap-rag hooks. How to use it, what it never records, how it was tested and what it does not do.
v0.13.0: documents are now untrusted content for actions, and Stress Office tests the guardrails (100 runs, a small test, not a certification)
ragleap-core v0.13.0: answers built from your documents now need owner approval for outbound actions, even in full autonomy. Stress Office, a small open guardrail test: what it shows and what it does not.
Community test: RagLeap Core on Windows 11 with Ollama, what was shown and what was not
One community tester ran the RagLeap Core Windows installer with the Ollama option on one Windows 11 Pro machine. What the evidence shows, where it is, and what it does not prove.
ragleap-physical: let an AI propose, a person approve, and test it all in simulation first (pre-alpha)
How ragleap-physical, a pre-alpha simulation-only safety bridge, keeps the AI to proposals while a named person approves the exact change. Its limits, tests, Windows findings and how to help.
What people say about RagLeap: two public LinkedIn posts from people outside the maintainer team
Two public LinkedIn posts from people outside the RagLeap maintainer team, with links to the originals, the screenshots we highlighted, and what each post does and does not show.
v0.12.1 security release: the voice channel now needs a secret token in its address, plus single sign-on and offline install, with the limits the notes admit
ragleap-core v0.12.1, a security release: the voice channel now rejects connections without a secret token. What to change after upgrading, what else is in it, and the limits.
v0.12.0: folder auto-ingest, four roles with personal keys and an audit log, and the limits the release notes admit
ragleap-core v0.12.0: optional folder auto-ingest (off by default), four roles with personal keys, an audit log of who made each state-changing request, Meta WhatsApp, Google Calendar and HubSpot items, and the known limits.
RagLeap Core is a candidate corpus in the XERJ Corpus Hub (in licence review)
The XERJ Corpus Hub lists antonyrag/ragleap-core as a candidate corpus for AI coding agents. What the Hub is, what is listed, and its status: selected, in licence review, not installable yet.
v0.11.0: Documents tab, a non-root container and experimental computer control, the first release under the ragleap organisation
ragleap-core v0.11.0: the Documents tab, 28 file types and web pages by address, Proactive and Connectors tabs, offline language detection, a non-root container, one scheduler, DemonRoute, experimental computer control (off by default, not independently reviewed), and the limits the release notes admit.
v0.10.0 and v0.10.1: login throttling, a Setup tab and installer fixes after a first Windows run
ragleap-core v0.10.0 added login throttling, a dashboard Setup tab, a read-only skill for coding agents and safer installer checks. v0.10.1 fixed installer problems from the first real Windows run. What changed, and what is still unfinished.
RagLeap is listed in the HOL plugin registry: what the review involved
RagLeap was merged into the awesome-ai-plugins catalog and the HOL Registry. The scores from 62 to 94, the scanner findings, and what the listing does and does not mean.
v0.9.0: the AI Office dashboard, a new installer and pluggable embeddings
ragleap-core v0.9.0 adds the AI Office dashboard at /office, a one-command installer with a ragleap command, embedding providers beyond Gemini and an agent toolkit, plus the limits the release notes state openly.
RagLeap Core: 46 built-in AI employee roles, self-hosted and MIT licensed
The 46 built-in AI employee roles in ragleap-core v0.9.0, how to self-host it with Docker, and how the installer and embedding options work.
How RagLeap detects the language of very short queries
How ragleap-core improved short-query language detection with FastText plus a character n-gram fallback, and what its 32-query benchmark does and does not show.
Going all-in on open source
We're retiring ragleap.com's cloud product and putting 100% of our effort into ragleap-core as an open-source project — MIT licensed, self-hosted, no account required.
v0.7.0: the 9-pattern agentic build is complete
All 9 standard agentic patterns are now shipped in ragleap-core — reasoning, reflection, routing, sub-agents, memory, and tools — on top of approval-gated actions, usage budgets, and real security hardening.
A security incident, and what we did about it
While reviewing authentication before starting new work, we found ragleap-core's API had zero auth, several Docker ports were exposed to the public internet, and approval replies weren't checked against the actual owner. Here's exactly what we found, what we fixed, and what forensics showed.
More posts
How to use @ragleap/rag in Node.js: sanitize text, chunk it and rerank results (early TypeScript port)
A walkthrough of @ragleap/rag 0.2.0 for Node.js 22 and later, using only its README examples: sanitize and chunk text, optional reranking, and what has and has not been checked in this early port.
How to keep an AI agent from using another user’s MCP connection (ragleap-integrations 0.3.0)
ragleap-integrations 0.3.0 adds McpHub: one MCP connection per user or tenant, and the model never chooses the scope. What it does, how it was checked and its limits.
How to build an AI assistant for a law firm or legal practice with RagLeap Core, and what to check first
Using the legal_intake role and related roles in RagLeap Core: what the code does for a sensitive domain, what to upload, and what a law firm must check about confidentiality, advice and approval.
How to build an AI assistant for an online store with open-source RagLeap Core
Using the support, sales, returns_refunds_agent, warranty_claims_agent, inventory_agent and delivery_dispatch_agent roles in RagLeap Core for an online store: what to upload, how customers reach it, and what the connectors do not guarantee.
How to build an AI assistant for a school or training business with RagLeap Core
Using the tutoring_agent and admissions_agent roles in RagLeap Core for a school or training business: what to upload, how students reach it, and what to check about children and personal data.
How to build an AI assistant for an accounting, tax or finance business with RagLeap Core
Using the finance, collections_agent, tax_preparation_intake, insurance_agent and compliance_officer roles in RagLeap Core: how sensitive roles are treated, what to upload, and what to check first.
How to add long-term memory to an LLM app with SQLite and trust labels (ragleap-memory, alpha)
How to use ragleap-memory 0.1.0, an alpha SQLite store for tenant-scoped LLM memory: install, add and search entries, supersede a changed fact, trust labels, erase, and what is not available or audited.
How to add tenant scopes, a tamper-evident audit log and secret redaction to a Python AI app (ragleap-security, alpha)
How to use ragleap-security 0.1.0: tenant scopes, a hash-chained audit log, secrets that do not print themselves and input limits, with the threat model and what is not covered.
How to let an AI agent delegate work to sub-agents, with limits and approval rules (ragleap-agents 0.5.0)
The sub-agent features of ragleap-agents 0.5.0 from its changelog: the delegate tool, budgets and depth limits, tainted runs, what happens when a child needs approval, and what was not verified.
How to build an AI assistant for your business with open-source tools (any industry)
Set up RagLeap Core v0.11.0 for any business: install, choose your AI, upload documents, pick one of 46 roles, connect a channel, control autonomy, and what to check before real customers use it.
How to build an AI assistant for a real estate agency with open-source RagLeap Core
Using the real_estate_agent, property_management_agent and coworking_space_agent roles in RagLeap Core: what to upload, how enquiries arrive, and what to check first.
How to build an AI assistant for a restaurant, hotel or hospitality business with RagLeap Core
Using the hospitality_agent role and related roles in RagLeap Core: what to upload, how guests reach it, what the connectors do not cover, and what to check first.
How to build an AI assistant for a clinic or healthcare practice with RagLeap Core, and what to check first
The healthcare_intake role in RagLeap Core is treated as a sensitive domain. What the code does, what to upload, and what a clinic must check about data, approval, security and compliance.
How to build a RAG chatbot in Python with ragleap-rag, pgvector and Gemini
A step-by-step guide to a self-hosted RAG chatbot in Python with ragleap-rag 0.14.0: install, ingest a document, ask a question, choose a vector backend, and what is not verified.
How to add a knowledge graph to a RAG system with ragleap-graph and Neo4j
How to use ragleap-graph 0.9.0 with Neo4j: install, store a document, find documents by entity, search related entities, optional LLM extraction, and the limits to know.
How to build an AI agent in Python that asks for approval before it acts (ragleap-agents)
How to use ragleap-agents 0.5.0: a small agent loop with a step cap, any model, tool policies, and pause and resume approval gates, with what was tested with a real model and what was not.
Self-hosting an AI chatbot with RagLeap Core: what you run, what you need and the limits
A short guide to running RagLeap Core yourself: the installer, what the stack contains, which AI providers it can use, where your data goes, and what has not been verified.
Open-source AI chatbots for businesses in India: what RagLeap Core does today and what is still research
What the open-source RagLeap Core does today for multilingual businesses (channels, language detection, your choice of AI) and what is still research (Indian-language speech), with the limits stated.
Voice and business AI: what exists in RagLeap Core and what is only research
The voice channel in RagLeap Core, what the pre-alpha ragleap-voice package is and is not, what has been measured, and what is planned but not available.
Guides
26 guides across 17 packages — pick a package.