The post-rerank relevance floor rejects noise but passes coherent
off-topic questions, and it gets worse as the library grows. Adding
Gutenberg prose, the Python tutorial and MDN pages to the eval corpus
took the off-topic leak rate from 22% to 67% at the 0.62 floor.
Score cutoff does not fix it. The weakest correct top hit scores 0.666
while "write me a haiku about autumn leaves" pulls Walden at 0.705.
A cosine floor and a top-vs-median contrast test both cost recall
faster than they cut leaks.
- Add rag.relevanceCheck (default off). After the floor, the tasks
model, or the chat model when none is set, gives one grammar-constrained
on_topic verdict for the whole retrieved set. Off topic means no
context and no citations. Any failure keeps the chunks unjudged.
- Add a Settings > Models toggle recommending an 8B+ tasks model.
On llama3:8b the check cuts off-topic leaks from 67% to 6% for about
5 points of core recall@5 and ~0.6s per turn. On 3B models it rejects
most valid questions, which is why it is opt-in.
- Add 6 distractor documents and an off_topic suite (18 unanswerable
questions, 7 distractor controls) to the eval corpus.
- Report per-tag leak rates and leaked cases with top scores in
eval:retrieval. Add --judge-model to measure the check and --dump to
export pre-floor candidate pools for offline gate simulation.
- Pin the check off in eval:generation so a local setting cannot move
scores.
- Record the distractor sweep on RAG_MIN_FINAL_SCORE. The 0.62 default
stays: 0.65 halves leaks at no measured recall cost but leaves only
0.016 of margin.
The corpus fingerprint changes to b4ebe5dce8699f5d. Retrieval reports
from before this commit are not comparable, and a new baseline is
committed.
Refs #1341
Chat sent num_ctx, num_predict and nothing else, so every turn decoded at
whatever the backend picked: temperature 0.8, top_p 0.9, no min_p. On a
heavily quantized 1-3B model that is the wandering, repetitive output people
read as "the local model is bad".
Adds `ai.responseStyle` with four profiles. min_p is the lever, not
temperature: it floors relative to the top token, so it truncates hard on a
confident distribution and barely at all on a flat one, which is where small
models start to wander. top_p stays at 1.0 rather than stacked on top of it.
repeat_penalty moves with the style, because a grounded answer legitimately
repeats a part number inside the 64-token window: `focused` relaxes it to
1.05, `creative` tightens it to 1.15.
`auto` is the only style that defers to the model. Plenty of library models
ship author-tuned samplers (qwen3 sets 0.6/0.95/20), so those win key by key
and NOMAD fills only the gap, which in practice is min_p. A style the user
picked by name overrides them, or it would silently do nothing on half
their models.
The compat transport drops min_p, so a profile carries a separate compatTopP.
Passing top_p 1.0 to a backend with no min_p behind it would sample the raw
distribution, which is worse than the 0.9 it replaced.
Resolved in the controller, so the grammar-constrained task calls keep their
temperature 0. An explicit temperature still outranks a profile's, which is
what keeps the eval harness deterministic.
`eval:generation --style` is new and load-bearing: temperature 0 makes min_p,
top_p and top_k inert, so a default run cannot see this change at all.
`--style=off` sends no samplers and no temperature, and is the control.
Verified against Ollama 0.32.9, min_p accepted. 22 unit tests.
There was one cutoff, Qdrant's score_threshold at 0.3, and nomic scores
unrelated text well above it, so every question retrieved *something*
(nonEmptyRateOnRefusal 1.0) and rag_context wasted four lines asking a 3B
model to "silently judge" relevance. That is a classification a number
should decide, not a micro model.
Adds `RAG_MIN_FINAL_SCORE`, a floor on the post-rerank score, applied before
source diversity. The floor judges relevance, the 0.85^n penalty judges
redundancy. Nothing clears it and nothing is injected. `score_threshold` stays
at 0.3 as the candidate net. Exposed as `rag.minRelevance` with a module-
scoped TTL cache invalidated from `updateSetting`.
The floor is calibrated and the harness was describing the wrong number:
aggregate() bucketed semanticScore, so its percentiles measured raw cosine
while the new knob filters the reranked score. Both axes now, and the gap
matters: recall falls at 0.60 on the semantic axis, 0.66 on the final one.
floor recall@5 precision@5 emptyOnAnswerable nonEmptyOnRefusal
0.00 0.991 0.221 0% 100%
0.62 0.991 0.524 0% 60%
0.66 0.991 0.695 0% 60%
0.67 0.972 0.704 1% 60%
Default 0.62, not the last-lossless 0.66: the refusal win is complete at
0.62 and 28 documents is a thin basis for spending margin. 0.60 is the
honest floor. Three of five refusal goldens are near-corpus by
construction, where declining is the generation tier's job.
Both eval tiers pin the floor to the constant and ignore the setting, for
the same reason the harness never sets skipRetrieval.
4 metrics improved against baseline, no regressions.
Chat was run through Ollama's OpenAI-compatible /v1 endpoint, which has no
context-size field, so the num_ctx we sent was silently discarded and the
entire NUM_CTX_LADDER was dead code. Every conversation ran at the backend
default and was truncated from the middle (where history and
retrieved context actually sit). Measured with identical ~7k-token prompts:
2,060 tokens processed against 7,049 on the native endpoint, and with the
needed fact in the discarded middle the old path hallucinated a plausible
wrong number.
Transport: add a native /api/chat path using the already-installed ollama
package, sending a real options object (num_ctx, num_predict, temperature,
seed) plus keep_alive. The OpenAI path is still used for LM Studio and llama.cpp,
minus the ineffective num_ctx and with include_usage so both paths return
prompt token counts. /api/show is now read for trained context, parameter
size, quantization and any modelfile num_ctx. The <think> stream normalizer
moves to app/utils/think_stream.ts so both transports share it; it gains a
flush() that fixes a dangling partial tag being dropped at end of stream.
Token accounting: replace chars/N with a structural segmenter scaled by a
per-model factor learned by EWMA from the prompt_eval_count every response
already returned and threw away. End-to-end estimator error is 4.6% against
18-20% for a fixed divisor — qwen2.5 converges to k=1.26 where llama3 sits
at 1.00, so no single global constant serves both.
Context window: resolved once per model from trained context, the exact KV
cost derived from GGUF metadata (head_count_kv, not head_count — GQA makes
them differ 4x), and available VRAM, then memoized. Memoization is
load-bearing: asking for a different num_ctx makes Ollama unload and reload
the model, stalling the turn and discarding the KV cache. The floor never
overrides the model's trained ceiling — tinyllama trains to 2048, and a
blind 4096 floor pushes RoPE past anything the weights saw.
Prompt budget: planPrompt allocates explicitly instead of delegating
overflow to silent backend truncation — response reserve, then fixed blocks
(system and current query, never dropped), then RAG chunks whole and
best-first, then history in whole user+assistant turns, evicted newest-first
in blocks with an elision marker. Block eviction keeps the prefix stable for
several turns rather than shifting on every one.
KV-cache ordering: the per-turn retrieved block moves behind history to the
tail, so the stable prefix stays byte-identical turn to turn. At turn 14 of
a long conversation, prefill is 231ms at the tail against 539ms at the front
(~1ms/turn growth against ~25ms). The win is conditional — with a large
retrieved block and short history the two measure the same — so both
orderings ship behind RAG_PLACEMENT. The per-turn query rewrite also moves
to the tasks model; on a single-slot Ollama it was evicting the chat model's
cached prefix on every single turn.
Also: OLLAMA_CONTEXT_LENGTH and OLLAMA_KV_CACHE_TYPE=q8_0 on the container
(q8_0 is treated as headroom, never as a budgeting assumption, since
unsupported architectures fall back to f16 silently), a Context Window
setting in Settings > Models, and budget fields on PipelineTrace.
Eval: golden suites (--suite) so the new long_context goldens cannot move
the aggregates the committed baselines were recorded against, plus budget
and estimator-error metrics through the report path. Retrieval is unchanged
and eval:compare reports no metric outside the tolerance band; the
long_context suite scores correctness 1.000 with historyElided 0.500 (the
two eviction cases, by design). Unit tests 337 -> 386 passing, same 7
pre-existing failures.
Ingest-side CHAR_TO_TOKEN_RATIO is deliberately untouched: changing it
re-chunks the corpus and invalidates retrieval baselines, which belongs in
its own measured change.
