c0ecf2fa3bebd46bdfd0169efd73b89653d1dfed
2
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
3829b6add3 |
fix(ai): complete plain-key self-hosting path (#1584)
* feat(ai): resolve the Claude backend from the environment Tier 1 of #1406: a self-hosted deployment can now run every AI feature on a plain ANTHROPIC_API_KEY, with no AWS account. Hosted behaviour is unchanged. lib/ai/provider.ts resolves the backend once, from the environment: AI_PROVIDER explicit override, bedrock|anthropic AWS static key pair Bedrock ANTHROPIC_API_KEY the direct Anthropic API nothing set Bedrock, so the AWS credential provider chain (instance profile, IRSA) still resolves Bedrock deliberately wins when both credential sets are present. EU residency in eu-north-1 is a BFL/GDPR posture rather than a default, so adding an Anthropic key for an experiment must not silently move production inference out of the region. AI_PROVIDER is the way to say you meant it. Model ids are written bare in code and prefixed to eu.anthropic.* only for Bedrock, which needs the cross-region inference profile for on-demand throughput. An operator override that already carries a prefix passes through untouched, so BEDROCK_MODEL_ID and friends keep working as written. Converted call sites: the agent composer, invoice-inbox extraction, the document-extraction model label, and both receipt-hunt clients. The last two are not named in the issue, which predates receipt-hunt landing in main. @anthropic-ai/sdk is declared at 0.95.0, the version @anthropic-ai/bedrock-sdk 0.29.1 already pulled in transitively, so the lockfile dedupes to one copy with no new download. scripts/smoke-bedrock.ts becomes scripts/smoke-ai.ts and grows two steps. Unit tests can only prove which provider and model id get resolved; they cannot prove the resulting request is one the backend accepts. The script now sends real traffic over all three shapes the app uses: a plain create, a streamed turn carrying adaptive thinking, an effort level, an hour-long cache breakpoint and a tool, and document extraction end to end when given a file. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Bjorn Bergenheim <29535152+bjornbergenheim@users.noreply.github.com> * docs(self-hosting): document the AI smoke test The script added alongside the provider split is what closes the #1406 acceptance criterion ("document extraction and the assistant both work"), so a self-hoster needs to know it exists. Covers both invocations and states that it exits non-zero, which is what makes it usable as a post-deploy check. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Bjorn Bergenheim <29535152+bjornbergenheim@users.noreply.github.com> * test(ai): split the smoke test's thinking probe from its tool probe The combined probe could not falsify what it claimed to. It asked a question that needs a tool call, so the tool was used and adaptive thinking correctly declined to reason about it: the zero thinking-block count that came back was uninformative rather than a signal. 2a keeps the tool and drops thinking. 2b asks a question with several dependent steps (reverse charge, then a partial deduction, then the affected boxes) so that a model honouring the parameter must reason, and reports the thinking text length as well as the block count, since display:"summarized" can yield blocks with empty text. The cached system prompt is also padded past the 1024-token minimum cacheable prefix. Below that the API caches nothing and reports no error, so the old probe's cache counters read zero whether or not caching worked. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Bjorn Bergenheim <29535152+bjornbergenheim@users.noreply.github.com> * fix(document-extraction): stop requiring AWS_REGION in the manifest The extension now needs one of two credential sets, AWS static keys or ANTHROPIC_API_KEY, and the manifest schema cannot express "one of". Since requiredEnvVars only drives a build-time warning and never gates anything, listing AWS_REGION told every self-hoster running the direct API to set a variable that has no effect for them. The description was also still promising Sonnet 4.6 via Bedrock specifically, which is no longer what the extension does. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Bjorn Bergenheim <29535152+bjornbergenheim@users.noreply.github.com> * fix(ai): read documentKind defensively in the smoke test The field arrived with the receipt-aware extraction work, so referencing it directly stops the script compiling against any checkout from before that landed. tsconfig includes **/*.ts and next.config does not disable type checking, so on such a checkout this failed the production build rather than just the script: caught while preparing a test branch for a self-hosted instance that had not synced yet. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Bjorn Bergenheim <29535152+bjornbergenheim@users.noreply.github.com> * fix(deps): restore the nested @swc/helpers entry in the lockfile Declaring @anthropic-ai/sdk with `npm install --package-lock-only` also pruned node_modules/next-intl/node_modules/@swc/helpers@0.5.23, an optional peer entry the local npm 11 considers redundant and the image's npm 10.9.8 does not. The result passed every local check and failed `npm ci` inside the Docker build, which is the only place the lockfile is actually enforced. The lockfile is now the previous one plus the single root dependency line, verified with `npm ci --dry-run`. @anthropic-ai/sdk needed nothing else: it was already in the tree as a transitive dependency of @anthropic-ai/bedrock-sdk. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Bjorn Bergenheim <29535152+bjornbergenheim@users.noreply.github.com> * Update DECISIONS.md Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> * Update Docker documentation for AI provider credentials Clarify the role of credentials in AI provider selection and document extraction requirements. * Update SELF-HOSTING.md with smoke-ai script details Clarify usage of smoke-ai script for credential checks and document extraction. * Improve error handling and logging in smoke-ai script * fix(ai): complete plain-key self-hosting path Signed-off-by: Emil <emilmattsson14@gmail.com> --------- Signed-off-by: Bjorn Bergenheim <29535152+bjornbergenheim@users.noreply.github.com> Signed-off-by: Emil <emilmattsson14@gmail.com> Co-authored-by: Bjorn Bergenheim <29535152+bjornbergenheim@users.noreply.github.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> |
||
|
|
88140e5774 |
feat(receipt-hunt): let a model settle the pairs the arithmetic cannot (#1498)
The weights in the matcher are chosen by hand, and two of them were
adjusted last week because they produced a wrong answer on real data. A
pair agreeing to within 1%, from a merchant the matcher recognised, still
scored 0.62 because a date had drifted. That says more about the
constants than about the receipt.
So the formula keeps what it is good at and hands over what it is not.
Above 0.80 it decides alone: an exact amount from a recognised merchant
needs no second opinion, and buying one for every pair would be latency
and cost for a verdict nobody doubts. Below 0.60 nothing is asked either,
because inviting a model to rescue a pair the evidence does not support
is how plausible wrong answers get made. Only the band between goes for
adjudication, which on a real ledger was five pairs against one the
formula had already settled.
The question is a yes or no with a reason, never a score. An earlier
design in this feature asked a model to rate its own certainty and it
anchored on round numbers, which is what the calibration literature
predicts. Judging concrete evidence and explaining the judgement is a
different task.
A verdict is checked, not trusted: a key nobody asked about is dropped, a
pair answered twice keeps its first answer, and a failed call accepts
nothing, leaving the run exactly where the arithmetic left it. The
formula's own score is stored unflattered next to the verdict, because
dressing it up would hide the uncertainty that sent the pair for a second
opinion in the first place, and agent_metadata records which instrument
decided.
Adjudication runs on a dry run too. A provkörning is meant to show what a
real run would propose, and skipping it would show a smaller, different
answer than the one that lands.
Measured on a real ledger: five uncertain pairs asked, two accepted with
reasons a human can check ("Samma leverantör och belopp stämmer inom
rimlig valutamarginal"), three rejected. Proposals went from one to three.
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|