docs: OpenRouter response caching guide + rendered architecture diagram (#144)

* docs(examples): document OpenRouter response caching via extra_headers

OpenRouter Response Caching (sending X-OpenRouter-Cache / -TTL headers on
compiler LLM calls) is achievable through the general extra_headers
passthrough — no dedicated config needed. Document it in the configuration
guide alongside the other litellm/extra_headers recipes.

Refs #39

* docs: replace ASCII architecture diagram with a rendered image

Swap the text-based diagram under "Architecture" for a rendered
illustration (assets/openkb-architecture.png) covering the same pipeline:
raw/ intake -> markitdown (short docs) / PageIndex (long PDFs) -> LLM
wiki compilation -> the wiki/ foundation -> query/chat, the Skill
Factory, and future generators.

Author: KylinMountain

README.md

@@ -119,34 +119,9 @@ LLM_API_KEY=your_llm_api_key
 
 ### Architecture
 
-```
-raw/                              You drop files here
- │
- ├─ Short docs ──→ markitdown ──→ LLM reads full text
- │                                     │
- ├─ Long PDFs ──→ PageIndex ────→ LLM reads document trees
- │                                     │
- │                                     ▼
- │                         Wiki Compilation (using LLM)
- │                                     │
- ▼                                     ▼
-wiki/                                  │            ← the foundation
- ├── index.md            Knowledge base overview
- ├── log.md              Operations timeline
- ├── AGENTS.md           Wiki schema (LLM instructions)
- ├── sources/            Full-text conversions
- ├── summaries/          Per-document summaries
- ├── concepts/           Cross-document synthesis
- ├── entities/           Specific named things (people, orgs, places, products)
- ├── explorations/       Saved query results
- └── reports/            Lint reports
-                                       │
-                ┌──────────────────────┼──────────────────────┐
-                ▼                      ▼                      ▼
-            query / chat         Skill Factory          (future)
-          (LLM answers from    (redistributable       ppt / podcast /
-            the wiki)           agent skills)           report / …
-```
+<div align="center">
+  <img src="assets/openkb-architecture.png" alt="OpenKB Architecture: from raw documents (markitdown / PageIndex) through LLM wiki compilation to the wiki/ foundation, powering query/chat, the Skill Factory, and future generators" width="900" />
+</div>
 
 ### Short vs Long Document Handling
 

assets/openkb-architecture.png

@@ -136,6 +136,27 @@ litellm:
     Copilot-Integration-Id: vscode-chat
 ```
 
+#### OpenRouter response caching
+
+When your `model` is an `openrouter/*` model, you can opt into OpenRouter's
+[Response Caching](https://openrouter.ai/docs/guides/features/response-caching):
+identical-payload requests come back in ~80–300 ms with **zero token billing**.
+That's a direct win on the compile-retry path (a failed `add` re-runs every
+summary/plan/concept call with the same prompts) and on repeated `lint` / dev
+iteration. Send the cache headers via `extra_headers`:
+
+```yaml
+model: openrouter/anthropic/claude-sonnet-4.5
+language: en
+extra_headers:                  # top-level, or nested under `litellm:` — both work
+  X-OpenRouter-Cache: "true"
+  X-OpenRouter-Cache-TTL: "600" # optional, 1–86400s (OpenRouter default 300)
+```
+
+It's opt-in by design: responses are stored on OpenRouter, so leave it off for
+zero-data-retention / regulated content. Only `openrouter/*` models read these
+headers; other providers ignore them.
+
 ---
 
 ## 3. API keys & providers