Claus Ibsen created CAMEL-27221:
-----------------------------------

             Summary: camel-jbang - pattern recipes as Markdown in the CLI, and 
recipe lookup for local models
                 Key: CAMEL-27221
                 URL: https://issues.apache.org/jira/browse/CAMEL-27221
             Project: Camel
          Issue Type: Improvement
          Components: camel-jbang
            Reporter: Claus Ibsen
             Fix For: 4.24.0


Follow-up of CAMEL-23774 (pattern-level recipes catalog, PR #30345). The 
recipes are aimed first at local models (14B-35B): the ticket came out of our 
local-model benchmark runs, where small models found single options in the 
catalog but got the composition wrong. This ticket covers how users and local 
models get to the recipes.

h3. 1. One Markdown renderer

Render a recipe as Markdown once, and use it everywhere:
* title and problem
* components and EIPs
* recommended options as a table (option, value, why)
* pitfalls as a list
* the route as a {{yaml}} code block, the transport variants as extra sections
* requirements and links to the examples

h3. 2. CLI

* {{camel catalog recipe}} lists the recipes (id, name, problem)
* {{camel catalog recipe <id>}} prints one recipe as Markdown; an option writes 
it to a {{.md}} file
* next to the existing {{camel catalog doc}}, {{camel catalog component}} 
commands

h3. 3. Camel CLI terminal UI

* a recipe browser like the example browser: the list (name, problem) on one 
side, the rendered Markdown in the doc viewer on the other
* one key copies the route into the project, or runs it (the "runnable from the 
terminal" part of CAMEL-23774)

h3. 4. Local models: recipe vs camel_catalog_sample

The two tools work at different levels: {{camel_catalog_sample}} shows how to 
write one named thing (an EIP, a component, a data format) and where it goes; a 
recipe solves a problem with a whole route built from several of those, with 
its pitfalls. Small models pick a tool by its description, and both 
descriptions now talk about routes and snippets, so they get mixed up.

Options, from cheapest to boldest:
# Separate the triggers in the descriptions: a recipe when the user describes a 
behaviour ("avoid duplicates", "retry then dead letter", "rate limit", "call 
several services and combine"); a sample when writing a named EIP or component, 
or after a schema error.
# Point between the tools in their answers (small models follow a hint in a 
tool answer far better than one in a description): {{camel_catalog_sample 
name=idempotentConsumer}} (also circuitBreaker, saga, throttle) ends with 
"recipe available: idempotent-consumer"; a recipe answer ends with "check 
options with camel_catalog_doc, validate with camel_validate_source".
# Fold the recipes into {{camel_catalog_sample}}, e.g. 
{{name=recipe:idempotent-consumer}}, with the recipe ids in its list. One core 
tool fewer, which matters for the small-model prompt budget, and small models 
do better with fewer tools.
# One line in the system prompt: "for an integration problem, look for a recipe 
first".

Proposed: 3 + 2, and the tool answer is the Markdown of point 1 (fewer tokens 
and easier for a small model than escaped JSON with the YAML in a string).

h3. Measure

Benchmark prompts phrased as a problem ("don't process the same order twice 
from Kafka") and as a named thing ("add an idempotentConsumer"): which tool the 
model calls, and whether the route validates. Compare with and without the 
recipes.

_Created by Claude Code on behalf of davsclaus_



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to