Skip to main content
Choose the credential-free source demo to explore existing evidence and projections. Choose the persistent stack to submit new evidence and run model extraction. The HTTP examples below require curl and jq. For a released native executable, follow binary installation. That default-feature binary runs the model-backed demo-wiki and needs provider credentials. The credential-free command below compiles a different feature set; it is not a runtime switch on the distributed binary.

​
Get the source

Both routes start from a checkout. With Git installed, run:
Run the remaining commands from this directory.

​
Try the demo first

The Living Wiki demo needs Rust 1.94 or newer, but no database service or model credentials:
The first source build can take several minutes. Once the server starts:
  1. Open the Editorial context projection URL printed in the terminal, or find Meerkat Recall Living Wiki architecture in the context inspector.
  2. Read the projected context and its Compilation receipt. The receipt records what was compiled; it is not a judgment that every claim is true.
  3. Open the evidence browser and read the architecture source. Its Technical details link opens the inspector.
You have completed the demo when you have inspected a projection, its receipt, and the source evidence. Stop it with Ctrl-C. Demo data is held in memory and disappears when the process stops. This credential-free build leaves the article wiki at /wiki empty. The full demo uses a configured model to write articles. Continue below if you want to ingest your own evidence; stop the demo first so port 3000 is available.

​
Start a persistent local stack

Install Docker with Compose, then build and load the matching patched server using the paired server recipe. Set RECALL_SURREAL_IMAGE to its actual local tag or digest; an upstream SurrealDB image lacks the required capability. Compose refuses to start without this explicit selection. Copy .env.example to .env. The default Compose configuration uses OpenAI embeddings and an Anthropic extraction model. Set OPENAI_API_KEY and ANTHROPIC_API_KEY in .env before starting; model calls can incur charges. Credentials do not create per-space processing approvals.
Migrations must run against the same namespace and database as the server. Running the migration inside the Compose service uses that service’s settings. Meerkat Recall runs managed workers in the server process by default. Create the named space and its approved query/document embedding, extraction and support verification bindings before semantic ingestion. See the service setup for the current named-space setup; environment keys alone do not make it ready.
Compose is a local development configuration with authentication bypass. Before exposing an instance, configure authentication and network access as described in Deployment.

​
Check readiness

Wait for semantic_ingest_ready to be true. If it remains false, inspect issues, provider configuration, and service logs. GET /v1/health checks service health; it does not substitute for semantic readiness. For MCP, connect using Streamable HTTP at http://localhost:3000/mcp and call get_runtime_status with {}. Ingest tools are hidden until the runtime is ready. See MCP profiles for tool availability.

​
Ingest synthetic evidence

The following example is fictional. With local development authentication, X-Space-Id selects the space. When authentication is enabled, send a Bearer token whose space_id and permissions authorize the request instead.
The response includes doc_id, source_id, and work_item_id. Acceptance means the work was queued, not that extraction has finished. To use MCP, call ingest_doc_simple with the same title and text.

​
Track processing

The MCP equivalent is get_doc_pipeline_status with {"id":"doc:YOUR_ID"}, available on /mcp/full to authorized callers. Check extraction_state before moving on:
  • For pending or running, repeat the status request after the workers have had time to process the document.
  • For failed, inspect and resolve the failed work before retrying it.
  • For not_triggered, check readiness and the work queue if the state persists after ingestion was accepted.
  • For completed, inspect assertions_count and the results below. This state describes tracked work, so it does not guarantee useful or correct claims.
See the state definitions for the exact meaning of each value. For semantic search, also check that chunks.embedded has caught up with chunks.total; extraction and embedding finish separately. The source-first flow stages metadata, renders immutable document evidence, plans extraction, validates and commits artifacts, and emits outbox events for truth maintenance. The exact runtime stages are documented in the Pipeline overview.

​
Search and inspect

Use k to request semantic document search; without it, this REST route uses its lexical search path. Semantic search can fall back to lexical retrieval when embeddings are unavailable.
On the default MCP profile, use search_all for discovery and get_entity for returned entity IDs. Extraction is model-driven, so inspect the assertions and evidence rather than assuming exact entity names or relationship counts. You have completed this walkthrough when the server is ingest-ready, the submitted document has a processing outcome, and you have inspected the resulting evidence and search results. For the bigger picture, continue with the knowledge model or the worked example.

​
Next steps

Ingesting data

Work with files, sources, URLs, and media-backed evidence.

Security

Configure space isolation, scopes, clearances, and subject restrictions.

HTTP API

Integrate application code with the REST surface.

MCP tools

Inspect the available schemas and agent workflows.