Resume import fails with Anthropic: “compiled grammar is too large” #24

Activity

1. ruchindwd

   added

   bugSomething isn't workingSomething isn't working

   on Oct 28, 2026
2. Jchinonso commented on Oct 29, 2026

   

   Jchinonso

   on Oct 29, 2026

   More actions

   Same issue here: Dossier 1.9.2 built from source with Node.js, Anthropic claude-sonnet-5. Profile import from a PDF failed with "The AI provider ran into a problem", and Test connection worked. No failed request showed up in the Console's Logs.

   I tried the first fix suggested above. If the request with output_config.format gets a 400, it retries once without the format and puts the schema in the system prompt with withJsonSchema. The answer still goes through the same zod validation. Import worked on the first try after that. The AI tests pass: yarn vitest run src/lib/ai, 154 passed.

   Patch to src/lib/ai/providers/anthropic.ts:

   diff --git a/src/lib/ai/providers/anthropic.ts b/src/lib/ai/providers/anthropic.ts
   index 8e45d13..105648b 100644
   --- a/src/lib/ai/providers/anthropic.ts
   +++ b/src/lib/ai/providers/anthropic.ts
   @@ -4,7 +4,7 @@ import Anthropic from "@anthropic-ai/sdk";
    import { zodOutputFormat } from "@anthropic-ai/sdk/helpers/zod";
    
    import { AiError } from "../errors";
   -import { BaseProvider, type ChatAnswer, type ChatCall, type Effort } from "../provider";
   +import { BaseProvider, type ChatAnswer, type ChatCall, type Effort, withJsonSchema } from "../provider";
    import { keyGuardedFetch } from "./guarded-fetch";
    
    /** Offered when the model list can't be fetched, e.g. before a key is saved. */
   @@ -44,17 +44,29 @@ export class AnthropicProvider extends BaseProvider {
      protected async chat(call: ChatCall): Promise<ChatAnswer> {
        if (!this.client) throw AiError.config("Add your Anthropic API key in Settings → AI provider first.");
        const effort = call.effort && (await this.effortLevels(this.client, call.model, call.signal)).has(call.effort) ? call.effort : undefined;
   -    const outputConfig = { ...(call.json ? { format: zodOutputFormat(call.json.schema) } : {}), ...(effort ? { effort } : {}) };
   -    const message = await this.client.messages.create(
   -      {
   -        model: call.model,
   -        max_tokens: call.maxTokens + THINKING_TOKENS,
   -        system: call.system,
   -        messages: call.messages,
   -        ...(Object.keys(outputConfig).length > 0 ? { output_config: outputConfig } : {}),
   -      },
   -      { signal: call.signal },
   -    );
   +    const send = (structured: boolean) => {
   +      const outputConfig = { ...(structured && call.json ? { format: zodOutputFormat(call.json.schema) } : {}), ...(effort ? { effort } : {}) };
   +      return this.client!.messages.create(
   +        {
   +          model: call.model,
   +          max_tokens: call.maxTokens + THINKING_TOKENS,
   +          system: !structured && call.json ? withJsonSchema(call.system, call.json.schema) : call.system,
   +          messages: call.messages,
   +          ...(Object.keys(outputConfig).length > 0 ? { output_config: outputConfig } : {}),
   +        },
   +        { signal: call.signal },
   +      );
   +    };
   +    let message: Anthropic.Message;
   +    try {
   +      message = await send(true);
   +    } catch (error) {
   +      // Structured outputs refuse schemas with many nullable fields (the
   +      // profile import has 32) with a 400; the schema then goes in the prompt
   +      // instead, and the answer is still validated against it.
   +      if (!call.json || !(error instanceof Anthropic.BadRequestError)) throw error;
   +      message = await send(false);
   +    }
        if (message.stop_reason === "refusal") {
          throw new AiError("provider", "The model declined to answer this one. Try rewording it, or try again.", "refusal");
        }

   It retries on any BadRequestError when a schema was sent. You might prefer to match only the "compiled grammar is too large" message, so other 400s fail fast.
3. AienTech

   added a commit that references this issue on Oct 29, 2026

   Dossier 1.10.0

   ...

   a1ab921
4. AienTech commented on Oct 29, 2026

   

   AienTech

   on Oct 29, 2026

   Owner

   More actions

   Thanks both! Both parts are fixed in 1.10.0:

   • Importing a resume: when Anthropic refuses the answer's shape as too complex, Dossier asks once more with that shape in the instructions, and remembers to for that model (from fix: put the schema in the prompt when Anthropic refuses it as too complex #16).
   • Etc/Unknown: TZ now defaults to UTC, and Dossier uses UTC whenever the server's time zone isn't one it knows. If you use compose.production.yml, take the new one.

   @Jchinonso, your #30 does the same as #16, so I'm closing it in favour of the release. Thanks for testing it by hand.
5. AienTech

   closed this as completedon Oct 29, 2026