docs: clarify self-hosted provider alternatives in README and Concepts

[stoichy]

Summary

• README Option 1 now states Anthropic + OpenAI is the default, not the only path, and links out for alternatives
• Move OpenAI-only / upgrading notes out of cloud Option 2 so they don’t look like cloud setup steps
• Concepts: add “Prefer a different provider setup?” covering OpenAI-only, Gemini (same OPENAI_* vars), other embeddings, and fully local

Fixes issue #14

Also clarified alternatives in robrain.dev docs

[adelinamart]

Thanks Radu — good change. The provider table is much clearer than the single
OpenAI-only paragraph it replaces, and splitting the four paths out is the right
call.

I checked every env var in the new Concepts section against the code and they
all line up: LLM_PROVIDER, OPENAI_BASE_URL, OPENAI_LLM_MODEL,
OPENAI_EMBEDDING_MODEL, OPENAI_EMBEDDING_DIMENSIONS, EMBEDDING_PROVIDER
with the Voyage/Cohere keys. The 1536 advice is a good catch specifically —
embed() truncates anything longer, so asking Gemini for 1536 natively avoids
throwing away half the vector.

Four things before this goes in:

1. This is open twice. #15 and #16 are identical (same tree hash). Could you
   close one?

2. gemini-3.6-flash — did you actually run the Gemini path end to end, or
   is this written from the pattern? I'd rather not publish a model name we
   haven't tried.

3. Related: our OpenAI path sends response_format: {type: "json_object"}
   and max_tokens (packages/shared/src/llm.ts). Extraction needs strict JSON
   back. If Google's compatibility layer handles those differently, the path
   would look fine and quietly produce nothing. One real extraction run against
   Gemini would settle it — happy to merge once you've confirmed.

4. Small thing: moving the upgrade note into <details> dropped
   "(pulls the new Perception image and applies startup DB migrations)". That
   parenthetical is the reason people re-run up at all — can you put it back?

One note for next time on index.html: it's a built file, so a five-word edit
shows up as the whole page being rewritten. Nothing wrong with what you did —
both copies inside it (the HTML and the RSC payload) were updated consistently,
which is what I'd check for. Just flagging it so we both know the diff size is
noise, not scope.

[stoichy]

Hey @adelinamart, fixed the issues, PTAL. If you'd like I can also merge the 2 commits because they solve the same issue.

1. PR docs: clarify self-hosted provider alternatives #16 is closed now.

2 and 3. Gemini model + json_object / max_tokens
Ran the OpenAI-compatible path against AI Studio with response_format: {type:"json_object"} and RoBrain’s extraction budget (max_tokens: 300). Here are the result for 2 different models:

• gemini-3.6-flash returns an incomplete JSON due to extraction budget (probably because of thinking tokens)
• gemini-3.5-flash-lite returns a full valid JSON (decision / rejected / rationale)

For the embeddings gemini-embedding-001 at 1536 works ok.
Therefore I Updated Concepts to gemini-3.5-flash-lite.

4. Restored "(pulls the new Perception image and applies startup DB migrations)" in the upgrading details.