.gitignore

@@ -3,6 +3,7 @@ __pycache__/
 *$py.class
 venv/
 .env
+dev/.env
 .idea/
 .vscode/
 .agent/

README.md

@@ -104,14 +104,44 @@ Configure via environment variables.
 
 ## Known Issues
 
+- [ ] Update README with all updates
+
 - [x] When asking a follow up question the previous output disappears
-- [ ] Parts of the UI are not theme aware resulting in a unpolished look when not using a dark theme
+
+- [x] Parts of the UI are not theme aware resulting in a unpolished look when not using a dark theme
+
 - [x] When SearXNG provides a info blob for a search it appears on top of the overview i.e. `Wikipedia` or `Linux` 
 
 For any issues not stated here please create an issue ticket on [Gitea](https://git.tysstech.com/TySS-Dev/ollama-ai-answers-searxng/issues) or [GitHub](https://github.com/TySP-Dev/ollama-ai-answers-searxng/issues) and add the `bug` tag.
 
 ## Roadmap
 
+### Dev Server
+
+- [x] Stream viewer — show tokens arriving in real time in the debug panel as they come out of Valkey, so you can see exactly what the LLM is generating chunk by chunk
+
+- [x] TF-IDF score visualizer — show a table of which URLs were fetched, their scores, and which chunks were selected for context
+
+- [ ] Intent detection display — show what intent was detected and which system prompt was used for each query
+
+- [ ] Saved queries — save/load test queries so you can quickly re-run the same set of searches after making changes to the plugin
+
+- [ ] A/B model comparison — run the same query against two different models simultaneously and show both responses side by side
+
+- [ ] Response time breakdown — show how long each phase took: SearXNG fetch, page fetching, TF-IDF scoring, LLM stream start, stream complete
+
+- [ ] Context inspector — show the full assembled context string that gets sent to the LLM, so you can see exactly what it's working with
+
+- [ ] Prompt viewer — show the full system prompt + user prompt that gets sent to Ollama
+
+- [ ] Export button — copy the full context + prompt + response as a JSON blob for bug reports
+
+- [ ] Per-intent system prompt editor — edit the system prompts for each intent type live without restarting
+
+- [ ] Token counter — show estimated token count of the context being sent
+
+### Plugin
+
 - [ ] Working on feature plans
 
 ## Architecture

ollama_answers.py

@@ -1725,6 +1725,16 @@ class SXNGPlugin(Plugin):
 
             job_id = hashlib.sha256(f"{time.time()}{q}".encode()).hexdigest()[:16]
 
+            # Persist intent for dev UI
+            logger.warning(f"INTENT BEFORE PERSIST: {repr(intent)}")
+            logger.warning(f"JOB_ID BEFORE PERSIST: {repr(job_id)}")
+            try:
+                vk = _get_valkey()
+                vk.setex(f"ai:job:{job_id}:intent", 3600, intent)
+                logger.debug(f"{PLUGIN_NAME}: persisted intent '{intent}' for job {job_id}")
+            except Exception:
+                logger.exception(f"{PLUGIN_NAME}: failed to persist intent")
+
             payload_dict = {
                 "model": effective_model,
                 "messages": [
@@ -2038,6 +2048,17 @@ class SXNGPlugin(Plugin):
             detected_intent = _detect_intent(q_clean)
             js_intent = safe_json(detected_intent)
 
+            # Persist intent for dev tooling / UI
+            try:
+                vk = _get_valkey()
+                vk.setex(
+                    f"ai:job:{job_id}:intent",
+                    1800,
+                    detected_intent
+                )
+            except Exception as e:
+                logger.debug(f"{PLUGIN_NAME}: failed to persist intent: {e}")
+
             b64_context = base64.b64encode(context_str.encode('utf-8')).decode('utf-8')
             total_context_count = self.context_deep_count + self.context_shallow_count
 

requirements.txt

@@ -1,3 +1,4 @@
 flask
 flask-babel
 certifi
+python-dotenv