Blog

Which id format should an LLM see in tool calls? We tested three

· by

TL;DR. We tested raw database ids, word handles such as ws-amber-otter and letter-digit handles such as ws-k7q2 on three models. With raw ids, one model passed a wrong id in 3 of 40 trials. With either handle format, no model did in 200 trials. With word handles, one model asked which item the user meant in 3 of 6 trials, against 6 of 6 with raw ids, in each of two same-round pairs. The format did not change how often models showed an id to the user: 21 of 124 replies with raw ids and 18 of 121 with word handles. The samples are small, and the change in asking appeared on one model of three.

A model that calls tools must say which item a tool acts on. It reads a list of items and copies an id into the tool call. Plori runs AI agents in cloud Workspaces, and its chat assistant passes the id of a Workspace or a task in its tool calls. We tested which id format to show the model.

What did we test?

The chat assistant gets a list of the user's Workspaces and notices about their tasks. We compared three formats for the ids in them:

  • Raw ids are the database ids. A Workspace has a 36-character UUID, and a task has a 32-character hexadecimal string.
  • Word handles are a prefix and two words, such as ws-amber-otter and run-polite-river.
  • Letter-digit handles are a prefix and four characters that alternate letter and digit, such as ws-k7q2.

A handle is valid only inside one user's conversation. The platform replaces it with the real id before the tool runs. It rejects a handle that it did not issue.

The test had 20 cases and three models from three providers: model A, model B and model C. A full round had 200 trials: 6 per case for model A and 2 per case for models B and C. Code checked the tool name and the id argument. We ran the test on 2026-11-07.

Do models copy raw UUIDs wrong?

One model did. With raw ids, model B passed a wrong id in 3 of 40 trials. One was a UUID with six groups where a UUID has five. One had 13 characters in its last group where a UUID has 12. The third did not match the expected task id. Models A and C passed no wrong id.

Word handles gave 0 wrong ids in 200 trials, after we fixed a defect in the test harness. Letter-digit handles also gave 0 wrong ids in 200 trials.

Does the id format change whether the model asks which item the user meant?

On one model, yes. In this case the user says "update the report". Two Workspaces fit, "Sales report" and "Annual report", and "Sales report" is the one used most recently. The expected response is to ask which report.

We ran model C on this case in three pairs. Both arms of a pair ran in the same round, with 6 trials per arm. The table shows how often the model asked.

Pair Raw ids Word handles Letter-digit handles
1 6 of 6 3 of 6 not run
2 6 of 6 3 of 6 not run
3 4 of 6 not run 5 of 6

In pair 2, the instructions to the model were byte-identical in both arms.

Across all rounds, model C did not ask in 15 trials: 2 with raw ids, 11 with word handles and 2 with letter-digit handles. In all 15 it sent the task to "Sales report", the most recently used Workspace. Models A and B asked in every trial with every format.

We do not know why word handles made model C ask less often. The data shows only the rate.

Does the id format change whether models show ids to users?

No. With raw ids, 21 of 124 replies contained a task id. With word handles, 18 of 121 did. All of them answered a notice that a task had finished, failed or was waiting. One reply with a word handle was "Task run-polite-river has completed. The prices it reported are: ...". With raw ids, the same kind of reply gave the 32-character id.

Two other changes reduced the number. The notices named the task handle and did not name the Workspace. After we changed them to name the Workspace, 8 of 125 replies contained a task id or handle. Then we added a filter that removes a handle in parentheses from a reply before the user gets it. In the last round, 0 of 122 delivered replies contained one. We stored those replies after the filter, so we do not know how many the models wrote with a handle.

We first believed that models never showed raw ids. The raw-id round had no check for an id in the reply, and we read the missing check as zero. A search of the stored replies found the 21.

What did we ship, and what got worse?

We shipped letter-digit handles, the notices that name the Workspace and the filter.

Measurement Raw ids Word handles Letter-digit handles
Trials with a wrong id in a tool call 3 of 200, all from model B 0 of 200 0 of 200
Model C asked on "update the report", all rounds pooled 18 of 20 11 of 22 12 of 14
Replies with a task id or handle 21 of 124 18 of 121, then 8 of 125 after the notices named the Workspace 0 of 122, after the filter

One case got worse. The conversation says "We are working in Shop", and then the user says "check prices every Monday". The assistant must send the task to the Shop Workspace. In two same-round pairs, model C did this in 6 of 6 trials with raw ids and in 5 of 6 with letter-digit handles. Over all rounds it was 14 of 14 against 10 of 14. In the 4 failures it asked a question first, for example "Which products should I check prices for every Monday?". We shipped the letter-digit format with this result known.

What should you do in your own system?

  1. Do not make the model copy long raw ids. Give it a short handle, and replace the handle with the real id on the server. The model that copied UUIDs wrong made no copy error with either handle format.
  2. Do not build handles from real words. One of our three models asked less often on an ambiguous request with them.
  3. Do not expect an id format to keep ids out of replies. Name what the item belongs to in the same message as its id, and check each reply before you send it.
  4. Run every check on the baseline too. We did not, and we believed a zero that we had not measured.
  5. Test the ambiguous case on every model that you serve. Two of our three models showed no change.

What are the limits of this test?

  • The samples are small. Each same-round pair had 6 trials per arm.
  • The change in asking appeared on model C only.
  • The pooled rows in the table mix rounds with different prompt wording and notice text. The same-round pairs are the stronger evidence.
  • The 0 of 122 belongs to the notices and the filter together with the format. It does not measure the format.
  • All cases come from one assistant with one set of tools.