# ADR-034 — Remove NL/smart-search (supersedes ADR-028 ×2, ADR-034-ollama, ADR-035)

**Date:** 2026-06-07
**Status:** Accepted
**Issue:** #772
**Supersedes:** ADR-028 (nl-search-ollama), ADR-028 (ollama-docker-compose-service), ADR-034 (ollama-production-deployment-and-keep-alive), ADR-035 (rule-based-nlp-service)

---

## Context

The natural-language search feature ("KI-Suche" / smart search) allowed users to enter
free-form queries like *"Was hat Walter an Emma im Krieg geschrieben?"* and have them
interpreted by an LLM into structured filters (persons, tags, date range, keywords).

The feature went through two major iterations:
1. **Ollama integration** (ADR-028): an `ollama` Docker service running a local LLM
   (llama3.2/gemma3) parsed queries via a JSON-mode prompt.
2. **Rule-based NLP service** (ADR-035): after Ollama proved too slow and unreliable on
   CPU-only hardware, a Python FastAPI microservice (`nlp-service`, port 8001) replaced
   it with deterministic regex + spaCy parsing plus a lightweight LLM call.

Both approaches shared the same fundamental problem: inference on the production server
(Hetzner Serverbörse, no GPU, 64 GB RAM, i7-6700) was too slow to be useful, with
typical query latencies of 10–30 seconds. Users got better and faster results from
the existing keyword search with date/person/tag filters.

## Decision

**Remove the NL search feature entirely.** The Python `nlp-service` microservice, the
Spring Boot `search/` package (`NlSearchController`, `NlQueryParserService`,
`RestClientNlpClient`, `NlSearchRateLimiter`, and all supporting classes), the frontend
NL search components (`SmartModeToggle`, `SmartSearchStatus`, `InterpretationChipRow`,
`DisambiguationPicker`), the related Docker Compose services, Prometheus scrape job,
Grafana dashboard, and all i18n keys are removed.

The existing structured search (FTS keyword + person/tag/date/directional filters) is
sufficient for the archive's current audience and search workload.

## Consequences

- **Capability removed:** users can no longer enter free-form natural-language queries.
  They must use the structured filter bar (keyword text box + person/tag/date/directional
  dropdowns). For documents where these filters are sufficient, there is no regression.
- **Operational simplification:** the Docker Compose stack loses two services
  (`nlp-service` and previously `ollama`/`ollama-model-init`). Memory budget on the
  production host is freed. No external model weights need to be kept warm.
- **Future reinstatement:** if a GPU-capable host becomes available, re-implementing
  server-side LLM inference would be straightforward given the clean separation of the
  `NlSearchController` entry point. However, this ADR deliberately avoids leaving dead
  infrastructure or stub code in place — start clean if and when that becomes viable.
- **No data or schema change:** only query/endpoint code and Docker services are removed.
  The `documents`, `persons`, and `tags` tables and their FTS indexes are untouched.