Develop

.env_sample

@@ -17,3 +17,8 @@ OWNER_STATE_OR_PROVINCE=Denmark
 OWNER_COUNTRY_NAME=DK
 
 MAX_CONTENT_LENGTH=2097152
+
+# the assistant's plan service image tag (ghcr.io/atomgraph/webalgebra-server-ldh); `local` for an image built from a REST-VKG checkout
+# WEBALGEBRA_VERSION=latest
+# the OpenAI model the assistant plans and writes queries with; gpt-4o when unset
+# OPENAI_MODEL=gpt-4o

.github/workflows/release.yml

@@ -161,6 +161,13 @@ jobs:
           # strip the literal prefix
           echo "FULL_VERSION=${RAW_REF#linkeddatahub-}" >> $GITHUB_ENV
 
+      # The CLI resolves linkeddatahub-rdf at its own version, which the tagged checkout holds. Maven
+      # Central does not: release.sh deploys the library after pushing the tag, and Central publishes
+      # it minutes later, while this job runs seconds after the push.
+      - name: Build the linkeddatahub-rdf library
+        run: mvn -B install
+        working-directory: rdf
+
       # release.sh keeps cli/pom.xml at the platform version, so the tagged commit already carries it
       # and the jar manifest ldh --version reads gets it from there
       - name: Build the ldh CLI

AGENTS.md

@@ -73,6 +73,12 @@ The dataspace exposes a **read-only SPARQL 1.1 Query** endpoint (advertised via
 
 Write portable, standard SPARQL: use explicit `GRAPH` patterns, no engine-specific extensions.
 
+**The default graph is empty.** Every document is a named graph whose name is the document's URL, so a triple pattern outside `GRAPH` matches nothing. Put the patterns inside `GRAPH ?g { ... }` to query across every document, or inside `GRAPH <document URL> { ... }` for one document; a subquery needs its own `GRAPH` as well. A query that returns no rows without a `GRAPH` clause is not evidence that the data is absent.
+
+**One document per resource, so a join crosses graphs.** A resource is described in the document about it (`foaf:primaryTopic`), and the resources it links to - an order's customer, a product's category, an employee's manager - are each described in their own document, in another graph. A pattern that follows such a link therefore needs a `GRAPH` of its own for each resource: `GRAPH ?o { ?order schema:customer ?customer } GRAPH ?c { ?customer schema:legalName ?name }`, never both triples in one `GRAPH ?g`, which asks for them in the same document and matches nothing. Only a resource's own companions - its address, its contact point, an order's line items - share its document. When in doubt, give every subject variable its own graph variable: it costs nothing when two resources do share a document, and it is the only pattern that works when they do not.
+
+**A property is used on instances of its domain.** The ontology (`Link rel=lds:ontology`) declares each property's `rdfs:domain` and `rdfs:range`, and the data follows it: a property declared on `schema:Person` is not on the `schema:Corporation` that employs the person. To reach it, follow the ontology's path - the corporation's `schema:employee`, then the person's `schema:address`.
+
 Outbound `SERVICE` and `LOAD` — from the triplestore and from the platform alike — are routed through the `egress` forward proxy, which refuses loopback, private and link-local destinations. A federated query reaches public endpoints and cannot reach the deployment's own services. With no proxy configured and `ALLOW_INTERNAL_URLS` unset, in-JVM `SERVICE` (in a `PATCH` update or an import mapping) is disabled outright.
 
 ## Content & document model

CHANGELOG.md

@@ -1,3 +1,10 @@
+## [Unreleased]
+### Added
+- An assistant for a signed-in reader: a request in natural language becomes a Web-Algebra plan, shown with its operations and its XML, and runs only on Execute; the steps, the documents the plan wrote and its result are reported as they happen, and the result is answered in words and drawn as a table, a list, a grid or a chart
+- A conversation is an `ldh:Chat` block, placed in the document by an object block: the create bar's Assistant button starts one, each ends in its own composer, and every turn is stored as an `ldh:ChatTurn` with its plan and what the execution reported, so a chat is there when the reader comes back
+- A `web-algebra` service in the stack (`ghcr.io/atomgraph/webalgebra-server-ldh`), reached through nginx at the reserved `/webalgebra` path with a WebID certificate on the connection; it acts for the reader through the secretary agent's delegation, so a plan may write exactly what its reader may. Needs the `openai_api_key` secret
+- The agent guide (`AGENTS.md`) is served at `/AGENTS.md` on every dataspace origin, public and outside the dispatcher, so a client writing a query against the endpoint - the assistant's `SPARQLString` reads it as service documentation - learns that every document is a named graph and the default graph is empty
+
 ## [6.1.0] - 2026-10-05
 ### Migration
 - **BREAKING**: `ldh` drops `-b`/`--base`; dataspace-level commands take the base URI as their positional, defaulting to `LDH_BASE`

CLAUDE.md

@@ -108,6 +108,7 @@ The application runs as a multi-container setup:
 - **linkeddatahub**: Main Java application (Tomcat)
 - **fuseki**: One SPARQL server holding a TDB2 dataset per dataspace role (`config/fuseki/config.ttl`), named after the dataspace origin (deployment host dropped, role appended: `end-user`, `admin`, `northwind-traders.demo.end-user`, …), each under `fuseki/<dataset>/`; bound to apps in `config/system.trig`
 - **egress**: Squid forward proxy for the store's and platform's outbound requests (SPARQL `SERVICE`, `LOAD`): public destinations only, so a query cannot reach another dataset, Varnish or the platform
+- **web-algebra**: REST-VKG's `webalgebra-server-ldh` image — executes Web-Algebra plans and writes them from natural language for the assistant drawer (`xsl/client/chat.xsl`). nginx forwards `/webalgebra` (a reserved path, like `/static/` and `/uploads/`) to it with the caller's certificate; it presents the secretary's certificate to this instance with `On-Behalf-Of` naming the caller, which `WebIDFilter` honours because every agent's WebID document carries `<secretary> acl:delegates <agent>`. It addresses the instance by its public origin, rewritten to `nginx:9443` (`WEBALGEBRA_PROXY_HOST`), the same trick the platform's own `ClientUriRewriteFilter` plays
 - **varnish-frontend/varnish-admin/varnish-end-user**: Caching layers (admin and end-user caches both front the single `fuseki`)
 
 ### Data Flow

Dockerfile

@@ -24,6 +24,9 @@ RUN mvn -B -Pstandalone dependency:go-offline
 
 COPY src /usr/src/platform/src
 
+# the agent guide the WAR serves at /AGENTS.md (pom.xml, maven-war-plugin webResources)
+COPY AGENTS.md /usr/src/platform/AGENTS.md
+
 RUN mvn -Pstandalone clean install
 # ==============================
 

README.md

@@ -67,6 +67,7 @@ The [`ldh` command line interface](#command-line-interface) is attached to every
      - `secrets/client_truststore_password.txt`
      - `secrets/owner_cert_password.txt`
      - `secrets/secretary_cert_password.txt`
+     - `secrets/openai_api_key.txt` — optional: an OpenAI API key for the assistant drawer, declared in `docker-compose.override.yml` like the OAuth secrets; without it the assistant runs plans but cannot write them
      The one you will need to remember in order to authenticate with LinkedDataHub using WebID client certificate is `owner_cert_password`.
   5. Launch the application services by running this from command line:
      ```shell
@@ -217,6 +218,8 @@ _:warning: Do not use blank nodes to identify applications or services. We recom
     <dd>Password of the secretary's WebID certificate</dd>
     <dt><code>client_truststore_password</code></dt>
     <dd>Password of the client truststore</dd>
+    <dt><code>openai_api_key</code></dt>
+    <dd>OpenAI API key the <code>web-algebra</code> service writes the assistant's plans with. Optional, declared in the override together with <code>OPENAI_API_KEY_FILE</code>; the service executes plans without it, and the drawer then reports that a plan could not be written</dd>
     <dt><code>google_client_id</code></dt>
     <dd>Google's OAuth client ID</dd>
     <dd>Login with Google authentication is enabled when this value is provided</dd>

cli/pom.xml

@@ -4,7 +4,7 @@
 
     <groupId>com.atomgraph</groupId>
     <artifactId>linkeddatahub-cli</artifactId>
-    <version>6.1.0</version>
+    <version>6.1.1-SNAPSHOT</version>
     <packaging>jar</packaging>
 
     <name>LinkedDataHub CLI</name>

docker-compose.yml

@@ -5,6 +5,8 @@ secrets:
     file: ./secrets/secretary_cert_password.txt
   client_truststore_password:
     file: ./secrets/client_truststore_password.txt
+  # openai_api_key:
+  #   file: ./secrets/openai_api_key.txt
   # google_client_id:
   #   file: ./secrets/google/client_id.txt
   # google_client_secret:
@@ -28,6 +30,8 @@ services:
     depends_on:
       linkeddatahub:
         condition: service_healthy
+      web-algebra:
+        condition: service_started # its upstream has to resolve when nginx starts
     ports:
       - ${HTTP_PORT}:8080 # allow Tomcat to do HTTP to HTTPS redirect
       - ${HTTPS_PORT}:8443 # HTTPS
@@ -149,6 +153,38 @@ services:
         target: /etc/squid/squid.conf
     expose:
       - 3128
+  web-algebra: # runs Web-Algebra plans against this instance, and writes them from natural language for the assistant drawer
+    image: ghcr.io/atomgraph/webalgebra-server-ldh:${WEBALGEBRA_VERSION:-latest}
+    mem_limit: 1024m
+    restart: on-failure
+    depends_on:
+      - egress
+    expose:
+      - 8080 # internal only: nginx is the sole client, and the server trusts the Client-Cert header only because nginx sets it
+    environment:
+      # the origin that receives an identity - this instance, with every dataspace subdomain under it. Its public name is
+      # not a route to it from inside the network, so requests for it go to nginx's client-cert listener under that name
+      - WEBALGEBRA_TRUSTED_ORIGIN=${PROTOCOL}://${HOST}:${HTTPS_PORT}
+      - WEBALGEBRA_PROXY_HOST=nginx
+      - WEBALGEBRA_PROXY_PORT=9443
+      # the secretary's WebID: the agent every user's WebID document already delegates (root-owner.trig.template, SignUp),
+      # so a plan submitted by a user runs under that user's own access control, not the secretary's
+      - WEBALGEBRA_KEYSTORE=/var/webalgebra/ssl/secretary/keystore.p12
+      - WEBALGEBRA_KEYSTORE_PASSWORD_FILE=/run/secrets/secretary_cert_password
+      - WEBALGEBRA_TRUSTSTORE=/var/webalgebra/ssl/server/server.crt # the self-signed server certificate, as PEM
+      # the model key is optional, like the OAuth secrets: a deployment that writes plans declares the openai_api_key
+      # secret and this variable in its override; without them the service runs plans and declines to write any
+      # - OPENAI_API_KEY_FILE=/run/secrets/openai_api_key
+      - OPENAI_MODEL=${OPENAI_MODEL:-gpt-4o}
+      # public destinations (SPARQL endpoints a plan names, the model's API) go through egress like the platform's own; nginx
+      # is reached directly, since egress refuses internal addresses
+      - JAVA_TOOL_OPTIONS=-Dhttp.proxyHost=${EGRESS_PROXY_HOST:-egress} -Dhttp.proxyPort=3128 -Dhttps.proxyHost=${EGRESS_PROXY_HOST:-egress} -Dhttps.proxyPort=3128 -Dhttp.nonProxyHosts=nginx
+    secrets:
+      - secretary_cert_password
+      # - openai_api_key
+    volumes:
+      - ./ssl/secretary:/var/webalgebra/ssl/secretary:ro
+      - ./ssl/server:/var/webalgebra/ssl/server:ro
   varnish-frontend:
     image: varnish:7.7.3
     user: root # otherwise varnish user does not have permissions to the mounted folder which is owner by root
@@ -200,6 +236,10 @@ configs:
       }
 
       http {
+          upstream web-algebra {
+              server web-algebra:8080;
+          }
+
           upstream linkeddatahub {
               server ${NGINX_UPSTREAM_SERVER:-varnish-frontend:6060};
           }
@@ -217,6 +257,7 @@ configs:
 
           limit_req_zone $$limit_key zone=linked_data:10m rate=15r/s;
           limit_req_zone $$limit_key zone=static_files:10m rate=20r/s;
+          limit_req_zone $$limit_key zone=web_algebra:10m rate=5r/s; # a plan is a model call; polling is what bursts
           limit_req_status 429;
 
           client_max_body_size ${MAX_CONTENT_LENGTH:-2097152};
@@ -231,6 +272,30 @@ configs:
               ssl_prefer_server_ciphers on;
               ssl_verify_client ${NGINX_SSL_VERIFY_CLIENT:-optional_no_ca};
 
+              # the assistant's plan service. A reserved path, like /static/ and /uploads/: prefix-matched ahead of the
+              # generic location and never rewritten, so /webalgebra, /webalgebra/plans and /webalgebra/{id}/result reach the
+              # sidecar as they are. Reachable only with a WebID certificate on the connection: the sidecar acts for the
+              # WebID in the certificate nginx forwards, and a caller without one has nobody to act for. On-Behalf-Of is
+              # cleared for the same reason Client-Cert is - identity comes from the connection, never from a header
+              location ^~ /webalgebra {
+                  if ($$ssl_client_verify = NONE) {
+                      return 401;
+                  }
+
+                  proxy_pass http://web-algebra;
+                  limit_req zone=web_algebra burst=10 nodelay;
+                  proxy_read_timeout 120s; # writing a plan is a 10-30 s model call
+
+                  proxy_set_header Host $$host;
+                  proxy_set_header X-Forwarded-Host $$host;
+                  proxy_set_header X-Forwarded-Proto $$scheme;
+                  proxy_set_header X-Forwarded-Port ${HTTPS_PORT};
+
+                  proxy_set_header Client-Cert '';
+                  proxy_set_header Client-Cert $$ssl_client_escaped_cert;
+                  proxy_set_header On-Behalf-Of '';
+              }
+
               location / {
                   add_header Access-Control-Allow-Origin "*" always;
                   add_header Access-Control-Allow-Methods "GET, POST, PUT, DELETE, PATCH, HEAD, OPTIONS" always;

pom.xml

@@ -3,7 +3,7 @@
 
     <groupId>com.atomgraph</groupId>
     <artifactId>linkeddatahub</artifactId>
-    <version>6.1.0</version>
+    <version>6.1.1-SNAPSHOT</version>
     <packaging>${packaging.type}</packaging>
 
     <name>AtomGraph LinkedDataHub</name>
@@ -46,7 +46,7 @@
         <url>https://github.com/AtomGraph/LinkedDataHub</url>
         <connection>scm:git:git://github.com/AtomGraph/LinkedDataHub.git</connection>
         <developerConnection>scm:git:git@github.com:AtomGraph/LinkedDataHub.git</developerConnection>
-        <tag>linkeddatahub-6.1.0</tag>
+        <tag>linkeddatahub-5.5.4</tag>
     </scm>
 
     <repositories>
@@ -169,13 +169,13 @@
         <dependency>
             <groupId>${project.groupId}</groupId>
             <artifactId>client</artifactId>
-            <version>6.0.3</version>
+            <version>6.0.4</version>
             <classifier>classes</classifier>
         </dependency>
         <dependency>
             <groupId>${project.groupId}</groupId>
             <artifactId>client</artifactId>
-            <version>6.0.3</version>
+            <version>6.0.4</version>
             <type>war</type>
         </dependency>
         <dependency>
@@ -399,6 +399,17 @@
                     <webappDirectory>${project.build.directory}/${build.warName}</webappDirectory>
                     <failOnMissingWebXml>true</failOnMissingWebXml>
                     <attachClasses>true</attachClasses>
+                    <!-- the agent guide is served at /AGENTS.md on every dataspace origin (web.xml maps it to the
+                         default servlet), so a query-writing client can read how the endpoint is laid out; one file,
+                         kept at the repository root where agents expect it, and shipped from there -->
+                    <webResources>
+                        <resource>
+                            <directory>${basedir}</directory>
+                            <includes>
+                                <include>AGENTS.md</include>
+                            </includes>
+                        </resource>
+                    </webResources>
                     <overlays>
                         <overlay>
                             <groupId>${project.groupId}</groupId>

rdf/pom.xml

@@ -4,7 +4,7 @@
 
     <groupId>com.atomgraph</groupId>
     <artifactId>linkeddatahub-rdf</artifactId>
-    <version>6.1.0</version>
+    <version>6.1.1-SNAPSHOT</version>
     <packaging>jar</packaging>
 
     <name>LinkedDataHub RDF</name>

rdf/src/main/java/com/atomgraph/linkeddatahub/rdf/vocabulary/LDH.java

@@ -45,6 +45,10 @@ private LDH() { }
     public static final Resource CSVImport = ResourceFactory.createResource(NS + "CSVImport");
     /** ldh:RDFImport class */
     public static final Resource RDFImport = ResourceFactory.createResource(NS + "RDFImport");
+    /** ldh:Chat class: a conversation with the assistant, its turns the rdf:_N members */
+    public static final Resource Chat = ResourceFactory.createResource(NS + "Chat");
+    /** ldh:ChatTurn class: one question and what came of it */
+    public static final Resource ChatTurn = ResourceFactory.createResource(NS + "ChatTurn");
     /** ldh:MissingPropertyValue constraint class */
     public static final Resource MissingPropertyValue = ResourceFactory.createResource(NS + "MissingPropertyValue");
     /** ldh:ChildrenView resource */
@@ -62,6 +66,16 @@ private LDH() { }
     public static final Property delimiter = ResourceFactory.createProperty(NS + "delimiter");
     /** ldh:chartType property */
     public static final Property chartType = ResourceFactory.createProperty(NS + "chartType");
+    /** ldh:question property: what a chat turn asked */
+    public static final Property question = ResourceFactory.createProperty(NS + "question");
+    /** ldh:answer property: the assistant's answer in words */
+    public static final Property answer = ResourceFactory.createProperty(NS + "answer");
+    /** ldh:plan property: the turn's Web-Algebra plan, an rdf:XMLLiteral */
+    public static final Property plan = ResourceFactory.createProperty(NS + "plan");
+    /** ldh:execution property: what running the plan reported, an rdf:XMLLiteral */
+    public static final Property execution = ResourceFactory.createProperty(NS + "execution");
+    /** ldh:outcome property: how the turn went, in one line */
+    public static final Property outcome = ResourceFactory.createProperty(NS + "outcome");
     /** ldh:categoryVarName property */
     public static final Property categoryVarName = ResourceFactory.createProperty(NS + "categoryVarName");
     /** ldh:seriesVarName property */