Repository navigation
Retrieval helpers work again with every provider, and reranking orders results
by the model's judgment.
Fixed
- Query planning, expansion, HyDE, reranking and verification call the
bulk-tier model through the configured LLM backend at the least reasoning it
allows: thinking off for local models, the lowest declared effort for cloud
models. The helper used to call LiteLLM directly. That way:- A local Gemma 4 12B, which thinks by default, spent the reranker's whole
500-token budget reasoning and returned nothing, so reranking never ran. - Claude Haiku 5.5 and GPT-6 Luna rejected its fixed
temperature=0.3, so
every helper silently fell back for Anthropic and OpenAI chats. - API keys were looked up under the environment-variable name, which the
keystore does not use, so keys in~/.synthesis/keys.yamlnever reached
these calls.
The output budget now comes from the model'sdefault_max_tokens, and a
failed helper call logs the model and the cause.
- A local Gemma 4 12B, which thinks by default, spent the reranker's whole
- Reranking orders results by the LLM's judgment. Retrieval scores differ in
scale by source: cosine similarity for vector hits, a full-text rank that can
exceed 8 for keyword hits, and an RRF score near 0.03 for fused hits. Mixing a
raw score with the 0-10 judgment let keyword hits the model scored 0 outrank
the chunk it scored 10. On a personal workspace this buried the one relevant
note at eighth place, and the chat answered that it had no record. Now:- Retrieval strength is scaled to 0-1 before the mix.
- Results past the top 20 follow in retrieval order instead of re-sorting
above the reranked ones on raw scores. - A skipped rerank keeps the retrieval order.
- The context labels each chunk with the combined 0-1 score instead of its raw
score.
- Hybrid search fuses with RRF whenever either list has hits. It used to
return raw scores when one list was empty or keyword search failed, and those
outranked every RRF score in the shared sort. The HyDE boost now applies to the
key results are sorted by. thinking_effort: offswitches thinking off for local models. It used to
send no parameter, which left a model that thinks by default thinking. It now
sendsreasoning_effort: none, which LiteLLM passes to Ollama as
think: false.- Gemma 4 entries in
engines.yamldeclare their hybrid thinking. Ollama
reports the capability for all four. 12B, 26B and 31B think by default; E4B
does not. The 31B flagship now defaults to medium effort, as other local
flagships with thinking do.