Skip to content

Releases

ragleap-rag 0.16.0

Choose a tag to compare

@antonyrag antonyrag released this 15 Nov 04:54
107d12d

Upgrade note

  • Tenant views (RagLeap.for_tenant) on the Redis, OpenSearch and LanceDB backends of ragleap-vectorstores need ragleap-vectorstores>=0.7.0; on 0.6.x those backends ignore the tenant filter or cannot filter by tenant_id.

Added

  • Opt-in multi-tenant scope: RagLeap.for_tenant(tenant_id) returns a view that tags every ingest with tenant_id, adds it to every search filter, prefixes session ids, and lists, deletes and updates only that tenant's documents. RagLeap(..., require_tenant=True) makes unscoped calls raise TenantRequiredError. Default behaviour is unchanged; documents ingested without a tenant tag are not visible through a view.

Fixed

  • PineconeBackend, QdrantBackend, MilvusBackend and WeaviateBackend: the stored chunk fields (document_id, document_name, chunk_index, and for Milvus id and vector) can no longer be overwritten by caller metadata.
  • MilvusBackend metadata filter: string values are escaped (a value containing a double quote could change the expression), keys must be plain identifiers, booleans are written as true and false, and unsupported value types raise ValueError.

Changed

  • RagLeap.for_tenant: tenant ids are limited to 1 to 128 characters from A-Z a-z 0-9 _ . @ -, because the id is put into filter expressions.

ragleap-core v0.13.3

Choose a tag to compare

@antonyrag antonyrag released this 14 Nov 01:43
621fbf6

ragleap-core v0.13.3

File permissions for agents (behind switches, all off by default), a signed and attested image, and dependency fixes. Upgrading switches nothing new on.

What changed in the app

  • Agents can propose writing and deleting files, always with your approval. New tools write_file (#1390) and delete_file (#1397) work inside folders you grant with write access. The approval message and the dashboard show the real change (a diff, or the file's name, size and start). After approval exactly that change is made, and only if the file still matches what was previewed. Old versions are kept in .ragleap-trash inside the folder; you list, restore or empty them on the machine with ragleap files trash list|restore|empty (#1396). The agent cannot reach the trash. Approval is needed also in full autonomy mode, and the file text never goes into the action log.
  • Everything above is off by default. It needs the runner switch COMPUTER_CONTROL, ragleap files writing on on the machine, FILE_TOOLS_ENABLED and FILE_WRITE_ENABLED in the app (and FILE_DELETE_ENABLED for deleting), and a write grant. ragleap files add --mode write still refuses write grants on purpose.
  • Broad folders, short and read-only (#1398). A home folder, /opt, /mnt, /srv or /media can be granted with ragleap files add PATH --name N --broad --for 2h (1 minute to 24 hours, read-only). Names that start with a dot or look like secrets are hidden. This is a check on names, not on content.
  • Runner and preview groundwork: #1357 and #1378.

Supply chain and dependencies

  • Core image publishing (#1405). A release tag now builds the image with provenance and SBOM attestations, signs it by digest with cosign (keyless), runs a Trivy scan in report-only mode and attaches a CycloneDX SBOM of the pinned Python packages to the release. See docs/supply-chain.md. Honest status: this workflow had one manual test run; this is the first run on a real release tag.
  • Fixed versions in requirements.lock after the first image scan: soupsieve 4.0.1 (#1404), certifi 2026.10.18, oauthlib 4.0.1, pyasn1 0.6.5, pypdf 6.21.4 (#1406) and setuptools 81.0.0 (#1407, its bundled copies of jaraco.context and wheel were flagged).

Documentation

New guides: docs/file-access.md and docs/supply-chain.md.

Upgrade notes

  • No schema change, no new setting you must set. Rebuild and restart the app; the lock file changed, so the dependency layer is rebuilt (about 1 GB more free disk during the build). Pull a fresh base image when you build, so the system library fixes in it are included.
  • The sandbox runner changed. Rebuild the sandbox image only if you want to use file writing, deleting or broad folders.
  • The outside security review (#800) has not happened. Keep file writing off in production until you have reviewed it yourself. Nothing here makes an installation compliant with any law or standard, and a signature says who built the image, not that it has no vulnerabilities.

Thanks

Thanks to everyone who tested and reported, and to Paige Ortiz (CerTova AI) for the technical review that prompted the supply-chain work. Contributors are credited by name; tell us if you prefer not to be.

ragleap-agents 0.5.1

Choose a tag to compare

@antonyrag antonyrag released this 14 Nov 08:32
a79100c

ragleap-agents 0.5.1: tenant scope for runs, approvals and audit

Design: #1395. Code: PR #1411. Date fix: PR #1415. A security fix for shared multi-tenant use. Additive: a caller who passes no scope sees no change except the two items under Behaviour changes.

The gap this fixes (reproduced with the released 0.5.0 from PyPI)

One shared Agent and one shared state store. Tenant A's run paused for approval. Tenant B knew only the run id, called resume(run_id, {call_id: True}), and the tool executed with A's arguments. run() and resume() had no tenant argument, the store was keyed by run id alone, and the run id was 48 bits. Single-tenant use was not affected. If you run one agent for several tenants with 0.5.0, upgrade and pass a scope.

Behaviour changes (read first)

  • New run ids are 22-character URL-safe tokens (128 bits) instead of 12 hex characters. Code that assumed 12 hex characters breaks. Old ids still load and resume.
  • New dependency: ragleap-security>=0.2.0 (standard library only, same project).
  • RunContext has a new scope field (default None). With a scope given, a damaged stored state raises "unknown run" like an unknown run.

What is new

  • run(task, *, scope=None, tainted=False) and resume(run_id, decisions, *, scope=None). A run saves its scope; resume only continues it for exactly the same scope, every part. A wrong scope, an unknown run and a damaged state raise the same ResumeError("unknown run"). A run made with a scope cannot be resumed without one and the other way round.
  • Agent(..., require_scope=True), audit=AuditLog(...) (ids, tool names and sizes, never text), rate_limiter=RateLimiter(...) (per scope, before any model call). Sub-agents inherit the parent's scope.

What this does not defend

  • The caller must pass the right scope; the library does not log users in. A tool that reads another tenant's data by itself is not stopped.
  • A shared store file shows every tenant to anyone with access to the file; use one store per tenant for stronger separation.
  • An in-memory rate limiter is lost on restart. No independent security review has happened.

Verified (real output)

  • GitHub CI: 55 of 55 checks passed on the code PR and on the date-fix PR.
  • VPS (Python 3.10.12, SQLite 3.37.2, ragleap-security 0.2.0 and ragleap-tools 0.5.1 from PyPI): 251 tests passed. Sandbox: 251 on Python 3.12.3 and 3.10.20.
  • Mutation check (new script, covers only the 0.5.1 changes): 34 of 34 killed on Python 3.12.3 and 3.10.20.
  • From PyPI: the wheel's 6 Python files are identical to the source at the tag; the cross-tenant attack is refused with the published wheel; a paused run written by the real released 0.5.0 still resumes without a scope and is refused with one; the 251 tests from the source archive pass against the installed wheel.

Not verified

  • Windows and macOS. A real model provider (all tests use a scripted model). Store files on network file systems. The mutation run on the VPS.

Provenance

  • Tag ragleap-agents-v0.5.1 on a79100c. Uploaded to PyPI 2026-11-14T08:29:53 UTC by the Release workflow (environment pypi).
  • Wheel sha256 4b8c4def6df2923d1e2071e56b874da9f81a20d737642705e73c1a90e93113d4
  • Source archive sha256 88a4ac9428e263d255c0600a6bda7e7cda2a4192a97c2ec42e5623e7dcf61438
  • The provenance records of the wheel and of the source archive were decoded: repository ragleap/ragleap-core, workflow release.yml, environment pypi, digests identical to the files.

ragleap-vectorstores 0.7.0

Choose a tag to compare

@antonyrag antonyrag released this 13 Nov 05:12
f35845a

Added

  • Tenant filtering: tenant_id is now a filterable field in RedisBackend (TAG field), OpenSearchBackend (keyword, also mapped on existing indexes) and LanceDBBackend (nullable column, added to existing tables). Chunks without a tenant tag never match a tenant_id filter.

Changed

  • RedisBackend and OpenSearchBackend now raise ValueError for filter keys they cannot filter on. Before, Redis ignored them (so a filter such as {"tenant_id": ...} searched every tenant) and OpenSearch silently matched nothing. LanceDBBackend and UpstashBackend validate filter keys too.
  • A Redis index created before this change has no tenant_id field: filtering by tenant_id on it raises RuntimeError; use a new index_name.

Fixed

  • ChromaBackend.insert_chunk: caller metadata can no longer overwrite document_id, document_name, chunk_index or token_count.

ragleap-observability 0.8.5

Choose a tag to compare

@antonyrag antonyrag released this 13 Nov 06:07
dfc4dfa

Changed

  • Grafana image pinned to 12.4.15 (was 12.4.12). The 12.4.12 image carried one CRITICAL finding with a fix available, in its Alpine base layer (brotli-libs 1.2.0-r1, CVE-2026-111497, fixed in 1.2.1-r0; Trivy image scan of 2026-11-09). Trivy on 12.4.15 shows no fixable CRITICAL: the CRITICAL gate step passed on this image (workflow run 43199344988), and an earlier scan of the same tag (run 43024362905) printed 2 HIGH and 0 CRITICAL rows with a fix available. Chart install, upgrade and uninstall with the new image on Kubernetes v1.34.11, v1.36.4 and v1.37.0: all checks passed, including Grafana health (database ok), admin login, Prometheus and Loki datasource health, and no panic or fatal lines in the Grafana log, after install and after upgrade, with 0 level=error lines (workflow run 43199378315). Not verified: an in-place upgrade from an existing 12.4.12 database to 12.4.15 (the install test upgrades with the same image, so login on an existing database after the image change was not re-checked), and behaviour on a long-running cluster with real data.

Documentation

  • README, Availability SLO alerts: the 'Not verified' line was out of date. On a throwaway kind cluster with the real chart and a stand-in receiver (workflow obs-k8s-test, run 39112387967, 2026-10-18; the Kubernetes version was not pinned) the four rules loaded, Prometheus had an active AlertManager, a healthy system fired nothing, a stopped exporter fired RagleapAvailabilityFastBurn (190 s) and PostgresExporterDown (171 s), both reached the receiver and the resolved notification arrived, and a removed target fired RagleapTargetAbsent (581 s). Still not verified: slow-burn firing, a real Slack or email send, other Kubernetes versions, multi-node.

Verified on PyPI: the 0.8.5 wheel has 17 files under helm/; and the project Source URL is https://github.com/ragleap/ragleap-core/tree/main/packages/ragleap-observability

pip install ragleap-observability==0.8.5

ragleap-memory 0.3.1

Choose a tag to compare

@antonyrag antonyrag released this 12 Nov 10:15
c1a22ff

ragleap-memory 0.3.1: encrypted export and import

Design: #1294. Code: PR #1385. Date fix: PR #1386. Additive, no schema change. Encryption is done by ragleap-security (AES-256-GCM); this package only frames the file.

Behaviour changes (read first)

  • Plain export files (format version 1) and every call without the new arguments behave exactly as in 0.3.0.
  • Import refuses an encrypted file (header version 2) when no key or passphrase is given, and refuses a plain file when one is given (no downgrade).
  • A file whose header says version 2 is now read as an encrypted file. Before, it was refused as an unsupported version.
  • New extra: pip install ragleap-memory[crypto] (needs ragleap-security 0.2.0 or newer). Without it nothing else changes, and asking for encryption raises an error that names the extra.

What is new

  • export_file(scope, path, encrypt_with=None, passphrase=None, context=None); import_file and import_stream take decrypt_with, passphrase and context. A Key, a KeyRing (rotation) or a passphrase (scrypt, derived once per file).
  • Each line is bound to a fixed label, the SHA-256 of the exact header line, its position and your context. A wrong key, a wrong context, any changed byte, and a line that is moved, duplicated or taken from another file all fail with "line N: cannot decrypt". Dropped or extra lines fail the entry count.
  • Import stays the single safe path: the same trust cap, scope check, quotas, one transaction and limits apply after decrypting.

What this does not defend

  • It protects the export FILE, not the SQLite database, and not process memory while exporting or importing.
  • The header shows the number of entries, and line lengths show roughly how long each entry is.
  • Someone who holds the key can read and forge a file. An old file put back in place of a newer one is not noticed.
  • A decrypted file is still untrusted input. No independent security review has happened.

Verified (real output)

  • GitHub CI: 55 of 55 checks passed on the PR head (10aaff3) and on the date-fix PR (31d09c9).
  • VPS (Python 3.10.12, SQLite 3.37.2, cryptography 50.0.2): 532 tests passed. Sandbox: 532 on Python 3.12.3 and 3.10.20, and with cryptography 42.0.0.
  • Mutation check (sandbox): 232 of 232 killed on Python 3.12.3 and 3.10.20.
  • From PyPI: the wheel's 5 Python files are identical to the source at the tag; fresh installs work with and without [crypto]; the 532 tests from the source archive pass against the installed wheel; a database and an export file written by the real released 0.3.0 open and import; the released 0.3.0 refuses a 0.3.1 encrypted file with its unsupported-version message.

Not verified

  • The mutation run on the VPS. Windows and macOS. Several processes using one file; network file systems; big-endian machines.
  • Timing and memory with very large files: an import decrypts into memory up to max_bytes (64 MiB by default).

Provenance

  • Tag ragleap-memory-v0.3.1 on c1a22ff. Uploaded to PyPI 2026-11-12T10:09:29 UTC by the Release workflow (environment pypi).
  • Wheel sha256 a4057312ade8b3a1fa204a1e22d4a2cc07071eba4869934f3f5a7b4db0a5ee4f
  • Source archive sha256 8ecdb596ba0174737344489a49e14f93610c7ff183f030eee8a55f24199de691
  • The provenance records of the wheel and of the source archive were decoded: repository ragleap/ragleap-core, workflow release.yml, environment pypi, digests identical to the files.

ragleap-graph 0.11.0

Choose a tag to compare

@antonyrag antonyrag released this 12 Nov 04:27
9a6197c

Added

  • Temporal relation edges, opt-in: GraphIndex(temporal=True). A RELATES_AS edge now carries valid_from; when no document supports the relation any more it is closed (valid_to set) instead of deleted, and it is reopened when support returns (valid_from restarts, so an edge keeps one interval). With temporal=True, find_relations() and the community relation queries read only open edges. find_relations(..., as_of=<timezone-aware datetime>) returns the relations valid at that moment (edges without the properties count as always valid). Times are indexing times, not real-world times. With the default temporal=False every statement and result is unchanged; keep the same setting for writes and reads.

Fixed

  • Tenant isolation: document_entities, the community member-name lookup and search_related_entities now filter every entity or path node by namespace and user_id (they could return another tenant's entities only if an edge crossing tenants existed). New offline and live isolation tests.

ragleap-integrations v0.5.0

Choose a tag to compare

@antonyrag antonyrag released this 12 Nov 01:30
96fed9b

ragleap-integrations 0.5.0 (published to PyPI on 2026-11-11)

Two owner-controlled additions to the UTCP client, and one fix. Install: pip install ragleap-integrations==0.5.0 (optional extra: security). This text follows the CHANGELOG shipped with the release and adds the checks made after publishing.

Added

  • allowed_query on UtcpServerConfig and connect_utcp: the owner lists the static query parameter names a tool URL may carry (at most 8 names, each 1 to 32 of letters, digits, underscore, dot and hyphen, all distinct; anything else raises UtcpConfigError). Default empty: exactly the 0.4.0 behavior. Each pair in the URL must be name=value with a value of 1 to 64 of letters, digits, underscore, dot and hyphen; a repeated name, a name that is also an input property and any placeholder are refused. UtcpToolSpec.static_query keeps the pairs; they are sent first, then the arguments. An argument with the same name as a static parameter is refused at call time and nothing is sent.
  • One accepted auth shape: auth_type api_key, location header, var_name Authorization, api_key exactly Bearer followed by a ${NAME} placeholder, and no other key. It is accepted only because it describes what this client already does: the owner token as Authorization: Bearer to approved origins. NAME is never looked up, in the environment or anywhere. Without a token the tool is skipped (the template declares auth but no token is configured; hint: pass token). Any other auth block, and auth_tools, are skipped as in 0.4.0. A query parameter the owner did not allow gets the hint: list the name in allowed_query.
  • scripts/utcp_live.py (repository only): new option --allowed-query, and its description now says it makes two GET requests, not one. Design note docs/design/utcp-auth.md: the decisions and two corrections to the note were added.

Fixed

  • A tool URL ending in a bare question mark was accepted by 0.4.0 and a call went out with two question marks. It is now skipped (url has a query string or a fragment).

Verified

  • 626 tests (the 545 of 0.4.0 unchanged plus 81 new): 622 passed and 4 skipped without ragleap-security, 626 passed with it, on Python 3.10.12 in clean virtual environments. 48 of 52 deliberate breaks of the new code caught; the other 4 are equivalent mutants (another check on the same path gives the identical outcome). A first round of breaks found one real test gap, now closed.
  • After release (2026-11-11): the wheel and the sdist on PyPI have the same sha256 as on TestPyPI; a plain pip install fetches that wheel; the installed files equal the source; pip check is clean with and without the extra; a smoke test passes; the sdist tests pass against the installed wheel (Python 3.10.12: 622 passed and 4 skipped, or 626 passed with the extra). The sdist contains a copy of the repository .gitignore, added by the build tool and byte-identical to the repository one at the merge.

Verified after 0.4.0 (recorded here because published files cannot be edited)

  • 2026-11-07 15:39 UTC: the public manual at https://api.unified.to/utcp was read (HTTP 200, 33 tools, all http); 0.4.0 offered none of the 20 tried because every template declares auth. 2026-11-08 13:05 UTC: 35 tools with one uniform shape (POST, the auth block above, a format query, body_field arguments). 2026-11-09 03:19 UTC: 37 tools, each input schema one required object property named arguments. Counts and shapes only: no credential and no tool call.
  • Correction: the 0.4.0 notes said no public server with a current-format manual was found. That became wrong on 2026-11-07: one was found.

Not verified

  • Any tool call against a real UTCP server (discovery only was observed). The normative UTCP specification text. Whether a vendor returns the same manual when a bearer token is sent.

Known limitations

  • Not supported: other auth types (basic, OAuth2, other api_key shapes), header_fields, OpenAPI conversion, variable substitution and a multi-tenant UTCP hub. Tools whose input schema is larger than max_schema_chars are still skipped.

Release date note

  • The CHANGELOG heading says 2026-11-11 and the files reached TestPyPI and PyPI on 2026-11-11 (UTC).

ragleap-core v0.13.2

Choose a tag to compare

@antonyrag antonyrag released this 10 Nov 12:50
36aa5cf

ragleap-core v0.13.2

Audit and sign-in hardening. Everything new is read-only or opt-in, except the sign-in rate limit, which is on by default for the (off by default) single sign-on.

What changed in the app

  • Audit filters (#1306). GET /audit can be filtered and paged: method, path (starts with), status, since, until (ISO 8601) and before_id. Calls without them behave as before.
  • Audit retention (#1318), opt-in. AUDIT_RETENTION_DAYS, ACTION_LOG_RETENTION_DAYS and AGENT_RUN_RETENTION_DAYS (empty or 0 keeps everything, the default; 1 to 6 is refused). When a limit is set, the scheduler leader deletes older rows every 6 hours in small batches; runs that are running or waiting for approval are never deleted. Each purge writes a retention_purge AUDIT line. New routes: GET /retention (admin) and POST /retention/purge (owner, dry run by default).
  • Request id in the logs (#1319). The X-Request-ID of a call is now carried by the AUDIT log lines written while handling it (request_id). LOG_REQUEST_ID=true puts it in every app log line; off by default.
  • Rate limit on single sign-on (#1325). /auth/login and /auth/callback accept at most AUTH_SSO_MAX_PER_MINUTE (default 30, range 1 to 10000) requests per client and minute, then answer 429 (the callback redirects back with a message). The first refusal per client and minute writes a sso_rate_limited AUDIT line. Behind a reverse proxy list it in TRUSTED_PROXIES, otherwise everyone behind it shares one counter. AUTH_THROTTLE=off switches it off.

Documentation

New guides: docs/audit-retention.md and docs/audit-log-lines.md; docs/single-sign-on.md now describes the rate limit.

Upgrade notes

  • No schema change, no new dependency. Rebuild and restart the app.
  • All new settings are optional and default to the old behaviour, except the sign-in rate limit (on by default, generous).
  • Deleting history removes evidence: ship the AUDIT lines to a log store first and choose the retention period with your own counsel. Nothing here makes an installation compliant with any law or standard.
  • Not covered yet: the request id is not stored in the database rows, lines written later in the background have no request id, and stored prompts are not redacted.

Thanks

Thanks to everyone who tested and reported, and to Paige Ortiz (CerTova AI) for the technical review that prompted the audit work. Contributors are credited by name; tell us if you prefer not to be.

ragleap-core v0.13.1

Choose a tag to compare

@antonyrag antonyrag released this 09 Nov 12:35
5143283

ragleap-core v0.13.1

A small release. The only change in the app is a new structured audit log; everything else is documentation and CI.

What changed in the app

  • Structured audit log lines (#1278). Every agent action now also writes one JSON line on the ragleap.audit logger, starting with AUDIT : action_proposed, action_executed, action_error, action_rejected, action_skipped and action_forced_approval. Each line has a UTC time, the action type, channel, role, target (cut to 200 characters), who approved or rejected it (name, role, kind), an action_id that links a proposal to its decision, and a sha256 of the content and of the result. The text itself is never written. A log shipper can keep these lines.
  • Turn the lines off with AUDIT_LOG_LINES=false.

Upgrade notes

  • No schema change, no new route, no new dependency. Rebuild and restart the app.
  • The lines go to the app's log output only. How long they are kept depends on your own log setup. They do not make an installation compliant with any law or standard.
  • Not covered yet: GET reads, chat routes, per-step events of agent runs, the case where an approved action's send fails, and a retention or purge setting for the database tables.

Documentation

  • SECURITY.md refreshed: supported versions, scope, honest limits (#1276).
  • New operator guides: a security incident runbook (#1277) and a host time synchronisation guide (#1290).
  • GitHub Pages site fixed: the build, the README rendering and links that gave 404 (#1285, #1286, #1289).

CI

  • terraform init retries up to 3 times when GitHub returns a 504 (#1296).

Thanks

Thanks to Paige Ortiz (CerTova AI) for the technical review that prompted the audit and documentation work, and to everyone who tested and reported. Contributors are credited by name; tell us if you prefer not to be.