docs: add LLM inference in workflows section

API.md

@@ -783,3 +783,147 @@ curl -X POST https://api.riotpiao.com/workflow \
 - **Readiness:** `curl https://api.riotpiao.com/readyz`
 - **Models:** `curl https://api.riotpiao.com/v1/models`
 - **Logs:** `kubectl -n api logs deployment/homelab-frontend`
+
+---
+
+## LLM Inference in Workflows
+
+The Poimen workflows system includes built-in LLM inference activities that call `/v1/chat/completions` via the gateway.
+
+### LLMInferenceActivity
+
+Single-prompt LLM inference within a workflow.
+
+**Workflow Definition (Canvas Node):**
+```json
+{
+  "id": "llm-node-1",
+  "type": "llm-inference",
+  "label": "Analyze Code with LLM",
+  "data": {
+    "model": "reasoning",
+    "system_prompt": "You are a code analysis expert. Provide detailed feedback.",
+    "user_prompt": "Analyze this code for security issues: {{ previous_output.code }}",
+    "temperature": 0.7,
+    "max_tokens": 2048
+  }
+}
+```
+
+**Backend Implementation:**
+The LLMInferenceActivity in the workflows backend automatically:
+1. Substitutes template variables (e.g., `{{ previous_output.code }}`)
+2. Calls `/v1/chat/completions` with the resolved prompt
+3. Returns the LLM response as activity output
+4. Retries on transient failures (up to 3 attempts)
+5. Timeouts after 120 seconds
+
+**Output:**
+```json
+{
+  "response": "The code has several security vulnerabilities...",
+  "model": "reasoning",
+  "stop_reason": "stop_sequence",
+  "tokens_used": 450
+}
+```
+
+**Supported Models:**
+- `reasoning` — DeepSeek-R1-Distill (best for complex analysis)
+- `ornith:35b` — Ollama 35B
+- `ornith:13b` — Ollama 13B
+- `qwen2.5:3b` — Qwen 2.5 3B
+
+---
+
+### LLMBatchInferenceActivity
+
+Multiple-prompt LLM inference (sequential processing).
+
+**Workflow Definition:**
+```json
+{
+  "id": "llm-batch-1",
+  "type": "llm-batch-inference",
+  "label": "Batch Code Review",
+  "data": {
+    "model": "reasoning",
+    "system_prompt": "Review each code snippet and provide feedback.",
+    "prompts": [
+      "Review snippet 1: {{ files[0].content }}",
+      "Review snippet 2: {{ files[1].content }}",
+      "Review snippet 3: {{ files[2].content }}"
+    ]
+  }
+}
+```
+
+**Output:**
+```json
+{
+  "responses": [
+    "Snippet 1 review...",
+    "Snippet 2 review...",
+    "Snippet 3 review..."
+  ],
+  "model": "reasoning",
+  "errors": []
+}
+```
+
+**Typical Use Cases:**
+- Batch code review across multiple files
+- Parallel document summarization
+- Comparative analysis of alternatives
+- Policy compliance checking
+
+---
+
+### Workflow Integration Examples
+
+**1. Code Analysis Workflow**
+```
+Clone Repo → Analyze Code → LLM Security Review → Generate Report → Notify
+```
+
+**2. Document Processing**
+```
+Retrieve Documents → Embed + Index → LLM Summarize (batch) → Archive
+```
+
+**3. Multi-Stage Review**
+```
+Retrieve Memory → LLM Context Extraction → Route to Activity A/B/C → Notify
+```
+
+---
+
+### Error Handling
+
+If LLM inference fails:
+- First activity retry (2-second backoff)
+- Second activity retry (4-second backoff)
+- Third activity retry (8-second backoff)
+- If all retries fail, workflow records error and proceeds to next activity (or fails if terminal)
+
+**Common Failure Scenarios:**
+- Network timeout: `connection refused` (retry automatically)
+- Model not found: `unknown model: xyz` (terminal error)
+- Rate limited: HTTP 429 (retry with exponential backoff)
+- Prompt too long: `context length exceeded` (terminal error)
+
+---
+
+### Performance & Cost
+
+- Single prompt inference: ~100-500ms (model-dependent)
+- Batch processing: Serial (not parallel), ~100-500ms per prompt
+- Model inference costs: Free (on-premise Ollama/Reasoning models)
+- Token counting: Provided in response for quota tracking
+
+**Optimization Tips:**
+- Use `ornith:13b` or `qwen2.5:3b` for faster inference
+- Use `reasoning` only for complex analysis that needs reasoning
+- Cache frequently-used prompts at workflow level
+- Use batch activity for multiple similar prompts (better throughput)
+