Feature status
This page lists the capabilities of Foundation4 by status. The page is updated with each release and replaces the former Roadmap page. Available means the capability is reachable through the public API. Partial means the capability works within the limits stated in the notes. Planned means the capability is on the roadmap and not yet available.
Content and ingestion
| Feature | Status | Notes |
|---|---|---|
| Plain-text documents through the REST API | Available | Documents are submitted as JSON with a contents string |
| Asynchronous processing with status tracking | Available | The document status reports pending, success, failed or canceled. A failed document records the reason after 12 attempts |
| Document versions and updates by external identifier | Available | Previous versions are expired by default and remain readable |
| Point-in-time reads | Available | Returns a document, or the fragments of a document, as they existed at a specified time |
| Expiry and deletion | Available | Single documents, a whole pipeline, or all pipelines |
| Metadata with an optional JSON Schema | Available | Validated on ingestion |
| Text splitters | Available | Character, recursive character, code, Markdown header, token, NLTK, spaCy and sentence-transformers token splitters. The token splitter reads its encoding files from a model image or downloads them. The spaCy and sentence-transformers token splitters need their package images. The recursive JSON splitter cannot be created in the current release |
| Five embedding providers | Available | FastEmbed, OpenAI-compatible endpoints, GPT4All, Hugging Face sentence-transformers and Hugging Face inference endpoints |
| File upload and parsing (PDF, Office, HTML, audio) | Planned | Text must be extracted before submission |
Search and retrieval
| Feature | Status | Notes |
|---|---|---|
| Search indexes | Available | Vector, full-text and BM25 indexes per pipeline, added, rebuilt, switched or removed while the pipeline holds documents. Several embedding models per pipeline |
| Vector similarity search | Available | Cosine distance by default. Euclidean and dot product distance after adding a distance index |
| Maximal marginal relevance (MMR) search | Available | Balances relevance against variety |
| Full-text search | Available | Full-text index with PostgreSQL text search, language detected automatically across 36 languages, or BM25 index with a configurable tokenizer. Hugging Face tokenizers need internet access when the index is created |
| Document paths | Available | Searches, agents and document lists scoped to a path pattern. A search without a path covers only documents without a path. Paths restrict no access |
| History search | Available | The most recent fragments of a pipeline, newest first |
| Metadata filters | Available | Comparisons, ranges, pattern matching, set membership and boolean logic, on top-level and nested fields |
| Taxonomy-aware filters | Available | Match a value and every value beneath that value in a hierarchy |
| Custom metadata indexes | Available | Indexes on declared top-level metadata fields, built on the expression that metadata filters use. PostgreSQL decides for each query whether to use an index |
| Classification scoping | Available | Hierarchical or exact matching on every search, within the classification allow-list of the API key |
| Hybrid search | Available | One request that searches several indexes and fuses the results by reciprocal rank fusion, with equal weights and no threshold. See the guide Set up hybrid search |
Generation
| Feature | Status | Notes |
|---|---|---|
| Registered LLMs | Available | Any OpenAI-compatible model that the deployment can reach |
| Agents with prompt templates and retrieval placeholders | Available | Similarity, MMR, full-text, hybrid and history placeholders on the pipeline's default search indexes |
| Streaming responses | Available | NDJSON, server-sent events or plain text |
| Execution tracing | Available | Records the prompt and fragments behind a response for 60 minutes |
| OpenAI-compatible chat completions endpoint | Partial | Relays requests to a registered LLM so that OpenAI SDKs can call registered models. The endpoint does not add retrieval |
| Conversation history in agents | Planned | Agents are stateless; the client application supplies any history |
Agents, direct LLM queries and the chat completions endpoint call the chat completions API of the model server, so a model server that implements that API serves every feature. The API key of a registered LLM is optional.
Integration and access
| Feature | Status | Notes |
|---|---|---|
| REST API with OpenAPI description | Available | Served by every deployment at /openapi.json, with interactive documentation at /docs |
| API keys with read, write and execute permissions | Available | Per object type or per object, with expiry and deactivation |
| MCP server | Available | Streamable HTTP with 18 tools for pipelines, every search type, documents, taxonomies and agents, one prompt per agent and documents as resources. API key headers only, no OAuth authorization. Pipelines, agents and keys are created through the REST API |
| Administrative dashboard | Partial | Browse objects, add documents, run searches, test embedding models and LLMs. Pipelines, agents and API keys are created through the API |
Operations
| Feature | Status | Notes |
|---|---|---|
| Kubernetes installation with Helm | Available | Two charts: shared services and the application |
| Air-gapped deployment | Available | Model weights and packages ship as container images. The built-in FastEmbed models that are not in the API server image and the token splitter need internet access. The Models, packages and air-gapped installs page lists what works offline and provides the checklist |
| Prometheus metrics | Available | From the API server and the workers |
| OpenTelemetry tracing and request identifiers | Available | |
| Licensing | Available | License limits apply per deployment |