Skip to main content

Release notes

This page lists notable changes to Foundation4 since February 2026, newest first. Each heading names the month in which the changes under the heading were completed in the code.

November 2026 (release 1.1.0)​

Release 1.1.0 contains the changes completed in October and November 2026.

Search indexes. A pipeline holds any number of search indexes: vector indexes, one per embedding model, a full-text index and BM25 indexes. Indexes can be added, rebuilt, made the default or removed while the pipeline holds documents, and the workers fill a new index from the stored fragments without a new ingestion. A pipeline can therefore gain full-text search or move to another embedding model after creation. Ten operations under /pipelines/{id}/search-indexers manage the indexes. Search indexes describes the model.

Hybrid search. "type": "hybrid" searches the default vector index and the default text index, or any set of named indexes, and fuses the result lists by reciprocal rank fusion inside Foundation4. Set up hybrid search describes hybrid search.

BM25 ranking. A BM25 index ranks every fragment that shares a token with the query, with a tokenizer chosen per index: one of four built-in models or a Hugging Face tokenizer, with optional character filters, pre-tokenizers such as Jieba for Chinese, Snowball stemmers in 30 languages and stop word lists. Full-text searches read a full-text index or a BM25 index. BM25 tokenizers lists the options.

Document paths. A document version can carry a path, such as kb.billing.tax.vat. Searches, agents and document lists take a path pattern that selects the documents at a path, one level below a path or at any depth below a path, and GET /pipelines/{id}/paths lists the paths of a pipeline with their document counts. Document paths describes the model.

History search. "type": "history" returns the most recent fragments of a pipeline, newest first.

VectorChord. New vector indexes store vectors with the VectorChord extension of PostgreSQL. Vector indexes of existing pipelines stay on pgvector, and both return the same distances. Each distance strategy of a vector index has its own distance index, which can be added and removed.

Agent placeholders. Agents retrieve with full-text, hybrid and history placeholders as well as similarity and MMR placeholders.

Agent sources and attributions. An agent execution with "attributions": true records the answer in the execution trace and attributes each sentence of the answer to at most three retrieved fragments that support the sentence. The trace is read with GET /tracing/{id}. Over MCP, the run_agent tool returns the retrieved fragments in sources by default, and the attributions when the call sets attributions. Agents and prompt templates describes the attributions.

Failed and canceled documents. A worker retries a document that fails processing after a delay that starts at 10 seconds and doubles up to 10 minutes. After 12 failed attempts, about an hour, the document becomes failed with the reason in message. A version that is replaced or expired before processing becomes canceled. A document that fails no longer holds back the other documents of its batch. Ingest documents reliably describes the statuses.

Classification allow-lists. The classification allow-list of an API key applies to search, including the searches of agents and the MCP server, and to reads of documents and fragments, as well as to adding, expiring and deleting documents. A key searches and reads the classifications of the allow-list and the classifications that they inherit. Access control describes the rules.

MCP server. The MCP server offers 18 tools: pipeline discovery, every search type with paths and metadata filters, listing, reading, adding, expiring and deleting documents, taxonomies, agent retrieval and agent execution. Each agent is also an MCP prompt, and the documents of a pipeline are MCP resources. The MCP server keeps no session between requests, so any replica of the API server answers any MCP request. The MCP server accepts any host name by default, so MCP clients connect through the ingress as well as through a port forward. Connect an AI assistant over MCP describes the setup, and MCP tools lists the tools.

Administration. GET /license returns the license details, the limit of each object type and the number of objects in use. The response of POST /login reports in admin whether the key is an admin API key, such as the master key. GET /license and GET /statistics require an admin API key. API keys report in last_used_at the last time the key authenticated a request, recorded at most once an hour.

Cross-origin requests. The setting app.cors_allowed_origins, Helm value api-server.corsAllowedOrigins, lists the web origins that may call the API from a browser. API conventions describes the rules.

Metadata filters. One field takes several operators, which must all be satisfied, so a date range on one field needs no $and. Nested metadata fields are named with dots, such as author.name. Custom metadata indexes use the same expressions as filters. Filter operators describes the language.

List filters. An equality filter repeated in the query string of a list matches any of its values, as in GET /pipelines?name=a&name=b. Lists of embedding models filter by size, and lists of API keys filter by active.

Embedding models and text splitters. OpenAIEmbeddings takes dimensions, which asks an OpenAI-compatible model for shorter vectors. SpacyTextSplitter splits text into sentences with spaCy, and SentenceTransformersTokenTextSplitter splits text by the tokens of a sentence-transformers model; each provider needs its package image in the deployment. Fragments of MarkdownHeaderTextSplitter start with their heading lines. Providers and models lists the parameters.

Full-text languages. A full-text index records the language detected for each fragment and parses each query with the language of the fragment that the query is matched against, so short queries match regardless of the language detected for the query. Language detection recognizes 36 languages. Text detected as a language that PostgreSQL cannot stem, such as Ukrainian or Polish, is indexed without stemming, and Greek, Catalan, Hindi and several other languages use their own stemming and stop words where the database provides a text search configuration for the language. Full-text languages lists the languages.

Model servers. LLMs and embedding models accept endpoints on private, loopback and link-local addresses by default, where model servers inside the deployment's network usually run. The setting app.allow_private_endpoints: false limits endpoints to public addresses. An LLM can be registered without api_key for a model server that requires no key.

Request size. The API server accepts request bodies of up to 10 MiB by default, set with app.max_body_bytes. A larger request returns HTTP 413.

Changes for existing client applications.

  • Pipeline creation. full_text_search (fts or bm25) replaces has_full_text_search. A request that still sends has_full_text_search returns HTTP 422.
  • Pipeline response. default_vector_embedding_index_id, default_text_search_index_id, indexers and indexer_status replace default_vector_embedding_id, has_full_text_search and embedding_models. indexers maps each search index identifier to an object with type (vchord, pgvector, fts or bm25), the embedding_model of a vector index and the tokenizer of a BM25 index. PATCH /pipelines/{id} returns the pipeline with the new settings.
  • Unknown fields. Request bodies no longer ignore fields that the operation does not take. An unknown top-level field in a search request or in a request that adds a document returns HTTP 400 with a message that names the field, and the accepted fields in details.expected. An unknown field in most other bodies, including the create and update requests of pipelines, embedding models, text splitters, LLMs, agents, API keys and taxonomies, returns HTTP 422.
  • Search parameters. params.index names the index of a search, and a search that sends params.embedding or params.search_type returns HTTP 400. Any other key that the search type does not take, such as threshold in a hybrid search, returns HTTP 422. A parameter omitted inside params takes the default of the search type, where an earlier release used 0.
  • Distance strategies. A search with a distance strategy that has no distance index returns HTTP 400. New vector indexes have a distance index for cosine only.
  • Status codes. Full-text, hybrid and history searches no longer return HTTP 501. A search on a pipeline without the required index returns HTTP 400, and a search on an index that is still being built returns HTTP 409.
  • Errors. Server-side failures return HTTP 500 with the message Internal server error and no details, and the API server log records the cause. A failure of the model server during an agent execution or an LLM query returns HTTP 502, or ends a started stream with an error line, with the message Upstream LLM error. A failure to check an API key, such as an unreachable database, returns HTTP 503 rather than HTTP 401.
  • Error messages. A not-found response reads <Object type> not found, such as Pipeline not found, with the identifier in details. A permission refusal reads Unauthorized: <levels> access to <Object type> required, such as Unauthorized: read and execute access to Pipeline required, with the required levels in details.required. A request to an unknown path returns HTTP 404, and a method that a path does not accept returns HTTP 405, both with a JSON error body.
  • Fragments. Documents and search results carry a path field, which is null for a document without a path. Fragment lists and history search results carry no score field. Fragments requested by identifier with GET /pipelines/{id}/fragments carry their text and come back in the requested order, for up to 1,000 identifiers per request.
  • Classification allow-lists. A search that names a classification outside the allow-list of the key returns HTTP 403, and a key with an empty allow-list cannot search. A pipeline permission granted without classifications has no allow-list and permits no classification. Adding a document with a classification outside the allow-list also returns HTTP 403. Before an upgrade, each search key needs an allow-list that covers every classification that its requests name. For the Rocket.Chat integration, that is user and every role classification, as described in Rocket.Chat message search.
  • Document paths. A search, an agent request or a document list without path covers only the documents without a path, and no single request covers documents with and without a path. A client application that gives documents a path names a pattern in each request, and sends path with each new version, because a version does not inherit the path of the previous version.
  • Document statuses. Documents can end in failed or canceled as well as success. A client application that polls until success also stops on these statuses. A document whose processing job the queue does not confirm is returned by the create request with status failed.
  • Expired documents. GET /documents/{id} and the fragment list of a document return HTTP 404 for an expired document unless the request sets include_expired=true. Fragments requested by identifier leave out the fragments of expired versions unless the request sets include_expired=true. The version history of an unknown or deleted document returns HTTP 404 rather than an empty list.
  • Validation. Creating an embedding model or a text splitter with an unknown provider name returns HTTP 404 with Embedding Provider not found or Text Splitter Provider not found and the name in details.provider. A filter value outside the range of an integer field returns HTTP 400 with Schema error. A classification search_type other than 1, 2 or 3 returns HTTP 422. An agent whose placeholders are invalid is refused at creation and update with HTTP 400, Invalid placeholder <name>, the placeholder in details.placeholder and the reason in details.error.
  • Metadata filters. An empty $in list matches no document and an empty $not-in list matches every document, for fields of every type. An empty condition {} inside $or is ignored. A condition under $not never matches a document that lacks the filtered field.
  • Pagination. limit, first and last take values from 1 to 1000, and other values return HTTP 400 with Pagination error. $contains and $startswith treat % and _ as plain characters. order_by=+field sorts in ascending order without encoding the +. A cursor issued by an earlier release for a document list, or for a list with a custom order_by, returns HTTP 400 with Pagination error.
  • Cross-origin requests. The API server no longer accepts cross-origin requests from every origin. A web application that calls the API from a browser page on another origin needs the origin in app.cors_allowed_origins. The dashboard needs no entry.
  • Streaming. Streamed NDJSON responses carry the content type application/x-ndjson.
  • License checks. Agent executions and agent dry runs return HTTP 403 with Invalid license when the license is expired or over its limits, as direct LLM queries do. The MCP agent tools return JSON-RPC error -32602 Invalid license: <reason> in an HTTP 200 response.
  • MCP search tool. search_knowledge takes the search type in search_type, with similarity as the default and text for full-text search, the parameters as flat arguments such as k, threshold and fetch_k, and the search indexes in search_index_ids. The tool ignores the earlier arguments type, params and embedding_model_id, so a client that still sends them runs a similarity search with the default parameters. The pipeline objects of the MCP tools list the search indexes in search_indexers in place of embedding_model_id, with the kind vchord, pgvector, fts or bm25.
  • MCP sessions. The MCP server issues no Mcp-Session-Id and returns each response as JSON. GET and DELETE requests on /mcp return HTTP 405.
  • MCP errors. A tool call that the REST API would reject with HTTP 400, 403, 409 or 422, including an invalid identifier or a classification outside the allow-list, returns JSON-RPC error -32602 with the REST message, and a call for an object that does not exist returns -32002. Error -32603 reports only server-side failures, with the REST message, such as Internal server error or Provider unavailable, and run_agent reports a failure of the model server as The LLM call failed; see the server logs.
  • MCP agent execution. The run_agent tool is no longer annotated as read-only, so an MCP client can ask for confirmation before the tool runs.
  • Permission grants. A grant on contexts, context messages, documents, permissions or fragments, or on an object that does not exist, returns HTTP 400. A grant without classifications keeps the allow-list of existing entries.

Changes for operators.

  • Charts and image tags. The charts foundation4ai and foundation4ai-core are version 1.1.0. The charts set no default image tags: global.serverImageTag, global.grpcImageTag and global.dashboardImageTag, or the image tag values of each chart, name the image tags of the release, and a render without a tag fails with a message that names the value to set. global.foundation4aiImageRegistry sets the registry of the Foundation4 images, and pull secrets are given as plain names. The application chart requires Kubernetes 1.31 or later. Charts, images and installation bundle describes the values.
  • Secrets. The pods and the installation Jobs read the license, the application secret, the master key and the connection URLs as files under /app/config/secrets instead of environment variables. The core Secret holds only the connection URLs, the secrets file requires NATS_PASSWORD, and the passwords of the bundled database, cache and queue may contain any character. Secrets and keys describes the chain.
  • Restarts and installation Jobs. An upgrade of the application release restarts the pods when a Secret or the configuration changed, so no manual rollout restart is needed. The installation Jobs are deleted once all three succeed, and a failed Job remains with its log.
  • Bundled PostgreSQL. The bundled PostgreSQL of the evaluation profile uses an image with the VectorChord and BM25 extensions, creates the extensions at first start and keeps the database on a volume claim of 8 GiB, so the database survives a pod restart. Deleting the claim deletes the database. The initialization script of the 1.1.0 chart grants read access only to tokenizer_catalog, so creating a BM25 index fails with permission denied for table tokenizer until a superuser runs two grants once. Database describes the bundled database and the grants.
  • External PostgreSQL. An external database needs four extensions, preloaded with shared_preload_libraries and created by a superuser before the installation, and the database role must be able to call pg_control_system(), which computes the system ID, and needs write access to tokenizer_catalog for BM25 indexes. External database requirements lists the steps.
  • License. The license check Job fails at once on any result other than OK and keeps the system ID in its log. A version 3 license is bound to the PostgreSQL system identifier, the database name and the schema. When the license expires, the workers stop processing, and the documents queued meanwhile are processed if a renewed license is deployed within 24 hours. The document and fragment counts of the license limits are refreshed every 10 seconds. Licensing describes the binding and the enforcement.
  • Master key. At each installation and upgrade, the master key Job makes the stored master key match the configured identifier and secret, so both can be rotated. An expired, inactive or missing master key no longer stops the API server. Secrets and keys describes the rotation.
  • Health checks and probes. GET /healthz checks the database, the message queue, the cache and the gRPC provider service, and returns HTTP 503 when one of them fails. The charts use /healthz as the readiness probe of the API server and add a startup probe, so a slow database at startup does not restart the container. Health checks describes the probes.
  • Metrics. The API server serves /metrics on a separate metrics port, 9090 by default, set with app.metrics_port, and no longer on the API port. The bundled Prometheus server collects the API server and the workers with the scrape job foundation4ai-pods. foundation4ai_document_queue_processed counts every processing attempt, including retries, foundation4ai_document_queue_failed counts only the documents given up on, and foundation4ai_document_queue_processed_latency is a histogram. New metrics count retries, removed documents and agent executions and report the last activity of each worker. axum_http_requests_duration_seconds is a summary instead of a histogram, and requests to unknown paths share the endpoint label unmatched. Metrics lists the metrics.
  • Commands. foundation4ai data check-models checks that the configured models load, and replaces data download-models. foundation4ai admin keys-without-classifications lists the active keys that can search pipelines but name no classification. Command-line interface describes the commands.
  • Configuration. New configuration keys set the metrics port, the request size limit, the number of FastEmbed copies, private model endpoints, cross-origin origins, the database statement timeout and log level, the queue replicas and certificate authority, and the MCP host names, and prometheus.url sets the Prometheus server that /statistics reads. New chart values set resources and security contexts per container, PodDisruptionBudgets, the priority class and the startup probe. Configuration reference and Values reference list the keys and values.
  • Pod security. The API server, worker and installation Job pods run as the non-root user 10001 without privilege escalation or capabilities, which meets the restricted Pod Security Standard. Security hardening describes the settings.
  • Worker traces. With an OpenTelemetry endpoint set, the workers export traces and log records as the service foundation4ai-worker, linked to the trace of the request that added the document. Observability describes the export.

Upgrades. Existing pipelines keep their vectors in a pgvector index with the distance indexes for cosine, euclidean-distance and dot-product, and a pipeline that had full-text search receives a full-text index as the default text index. No content is ingested again. Fragments indexed for full-text search before the upgrade carry no language, so queries against those fragments are parsed with the language detected for the query until the documents are added again. The seeded default text splitter keeps its chunk size of 4,000 characters on an upgraded installation, while a new installation seeds 1,000 characters with 100 characters of overlap. The upgrade migrates the database, and the migrations cannot be reversed, so a rollback after the upgrade restores the database backup taken before it. After the upgrade, foundation4ai admin keys-without-classifications lists the keys whose searches find nothing because their allow-list is empty. An evaluation installation with the bundled PostgreSQL moves its database onto the new volume claim with a dump and restore, as described in Evaluation restore. Backup, restore and upgrades describes the procedure.

September 2026​

Full-text search. Pipelines can be searched by keyword as well as by meaning. A pipeline created with has_full_text_search: true accepts search requests with type: "search". Foundation4 uses PostgreSQL text search, detects the language of each fragment and each query automatically across 17 languages, matches all of the query terms, and ranks results by relevance. Full-text search applies classifications and metadata filters in the same way as vector search.

For hybrid results, the client application runs a vector search and a full-text search and fuses the two result lists. Release 1.1.0 adds hybrid search inside Foundation4.

Full-text search must be enabled when a pipeline is created. A full-text search request against a pipeline without full-text search enabled returns HTTP 501.

OpenTelemetry export. The API server exports a span for each API request, with the request identifier, and the log records of the API server through the OpenTelemetry Protocol (OTLP) when the environment of the API server sets an OTLP endpoint. The workers, the gRPC service and the dashboard exported no OpenTelemetry data; release 1.1.0 adds the export of the workers. Observability describes the export.

March 2026​

MCP endpoint. AI assistants that support the Model Context Protocol (MCP) can connect to /mcp over streamable HTTP, using the same API key headers as the REST API. Three read-only tools were available: list pipelines, list the classifications of a pipeline, and search a pipeline. MCP clients connected through a port forward to the API server. Release 1.1.0 extends the MCP server, as described under November 2026.

Verified database TLS. Connections to PostgreSQL can verify the server against a custom certificate authority.

Document processing metrics. The workers publish the time spent splitting text and computing embeddings as separate Prometheus metrics, next to the total processing time.

February 2026​

OpenAI-compatible chat completions endpoint. POST /openai/v1/chat/completions relays requests to a registered LLM, selected with the x-llm-id header, so that tools built on OpenAI SDKs can use models registered in Foundation4.

Embedding test endpoint. POST /embedding-models/{id}/query returns the embedding for a piece of text, which confirms that a model is configured correctly.

Branding. The product name shown in the API documentation and in the start log line of the API server can be set at installation, with the configuration key branding.name.

License limits. The API server checks the object limits of the license before each operation that creates or changes data.