Hybrid Pipelines Wikidata Agent

Flask API that exposes a single agent endpoint:

curl -X POST http://127.0.0.1:5050/analyze \
  -H "Content-Type: application/json" \
  -d "{\"text\":\"Mango is not a fruit from a tree.\"}"

The agent flow is intentionally simple:

1. Use the LLM to extract entities and concepts from the input text.
2. Use the configured Wikidata MCP server to search and inspect those entities.
3. Find direct Wikidata relationships between the resolved entities.
4. Ask the LLM to build RDF/Turtle from the text and Wikidata evidence.

Configuration

The Wikidata MCP server is configured with:

{
  "mcpServers": {
    "wikidata": {
      "type": "streamable_http",
      "url": "https://wd-mcp.wmcloud.org/mcp/"
    }
  }
}

The API uses the same endpoint through environment variables:

Variable	Default
SYSTEM_PROMPT_NAME	system/agent.txt
ENTITY_EXTRACTION_PROMPT_NAME	prompts/entity-extraction.txt
RDF_BUILD_PROMPT_NAME	prompts/rdf-build.txt
WIKIDATA_MCP_URL	https://wd-mcp.wmcloud.org/mcp/
WIKIDATA_LANGUAGE	en
WIKIDATA_TIMEOUT_SECONDS	60
WIKIDATA_CANDIDATE_LIMIT	3
WIKIDATA_ALLOW_ACTION_API_FALLBACK	true
WIKIDATA_USER_AGENT	hybrid-pipelines-agent/1.0
WIKIDATA_MAXLAG	5
WIKIDATA_MAX_RETRIES	2
WIKIDATA_RETRY_BACKOFF_SECONDS	2
OLLAMA_API_URL	http://localhost:11434
OLLAMA_MODEL	llama3:8b
OLLAMA_CSV_PATH	data/ollama_responses.csv
OLLAMA_TIMEOUT_SECONDS	300
ANALYZE_LOG_PATH	data/analyze_log.jsonl

Prompt files live under prompt/. Keep reusable task prompts in prompt/prompts/ and system prompts in prompt/system/.

Wikidata access follows the public access guidance: use Wikidata MCP for agent workflows, send a clear User-Agent, request gzip/deflate responses, pass maxlag to Action API fallback calls, and back off on 429 Too Many Requests.

Run

pip install -r requirements.txt
python -m src.app

The service listens on http://127.0.0.1:5050.

Response Shape

{
  "text": "Mango is not a fruit from a tree.",
  "entities": [],
  "relationships": [],
  "rdf": "@prefix ...",
  "llm": {}
}