Skip to content

README.md

Latest commit

 

History

History
65 lines (49 loc) · 3.44 KB

README.md

File metadata and controls

65 lines (49 loc) · 3.44 KB

Language & Translation Services (Transformers)

This directory contains language-oriented helper services that can plug into the broader OntoBot action pipelines. They provide:

  1. Deterministic (or model‑based) NL → SPARQL translation for ontology + telemetry joins.
  2. Local summarization / natural language post‑processing of analytics JSON and SPARQL results.

All runtime components in OntoBot are considered part of a flexible graph of pipelines; none is singled out as inherently “more core” than others. The Action Server can call these services when their corresponding features (translation or summarization) are desired.

Remote deployment note: Both nl2sparql and ollama can be hosted on separate machines or clusters. Simply set NL2SPARQL_URL and SUMMARIZATION_URL (e.g. to http://translator.host:6005/nl2sparql or https://llm.gateway/api) in the Action Server environment. If reachable over HTTP and following the expected request/response contract, the local containers for these services are not required.

NL2SPARQL (T5)

  • Purpose: Convert natural language questions into SPARQL queries.
  • Service: nl2sparql on port 6005 (compose + extras overlay).
  • Health: GET /health
  • Model: Transformers/t5_base/trained/checkpoint-3 (Download the files from Huggingface suhasdevmane/nl2sparql-brick-t5-base or search repository suhasdevmane/nl2sparql-brick on Huggingface)
  • Volume: Mount Transformers/t5_base/trained/checkpoint-3 read‑only into container (see compose)
  • ENV: MODEL_PATH=/app/checkpoint-3

If you omit this service, ensure the Action Server guards translation calls and either:

  1. Falls back to a rule/template set of SPARQL queries, or
  2. Routes the user toward intents that do not require dynamic SPARQL generation.

Ollama (Mistral)

  • Purpose: Summarization of SPARQL / analytics outputs, natural language refinement, and optional explanatory responses.
  • Service: ollama on port 11434
  • Health: root shows model list; use ollama ps inside container for process status
  • GPU: Supported (comment out GPU section if running on CPU only)
  • ENV recommendations:
    • AUTO_PULL_MODELS=mistral:latest
    • WARMUP_MODELS=true to pre‑generate tokens (reduces first‑response latency)

If disabled, set a flag (e.g., DISABLE_SUMMARIZATION=true) and ensure actions.py skips summarization stages gracefully (display raw metrics or structured JSON instead).

Customize / Extend

  • Swap checkpoints: adjust the mounted directory + MODEL_PATH env.
  • Multiple translators: add additional services (e.g., nl2sparql_large) and implement a switching rule in the Action Server.
  • Additional LLM runtimes: you can run a second ollama based model or integrate vLLM; keep the summarization contract stable.
  • Healthchecks: Always add a /health endpoint for orchestrator stability.

Feature Toggle Guidelines

If you run without one or both services:

Skipped Impact Mitigation
nl2sparql No automated SPARQL generation Provide template SPARQL or restrict intents
ollama No NL summaries; raw JSON/stat numbers shown Deterministic text assembly in Python

Document intentional feature toggles in the main README.md if you maintain a minimal deployment profile.

Quick Test Commands (PowerShell)

curl http://localhost:6005/health
curl http://localhost:11434

End of file