Files

Émile Ré 8182c61fa8 Improve common third-party enrichment quality

Give the company-profile agent (Agent A) the read-only browser toolset
and build it per-run, so it can read footer, imprint, about, and legal
pages and follow a product domain to the corporate one to resolve the
legal name and headquarters address rather than failing cold.

Make the website the hard precondition: when Agent A cannot resolve a
canonical website, skip the compliance-docs agent and logo step instead
of running them blind, which previously produced inconsistent
cross-domain document URLs. Fall back to the catalog display name for
the legal name when nothing better is found, recorded with a distinct
provenance status so a later real find overwrites it.

Rewrite both enrichment prompts in the project's role/task/instructions
XML style, add a domain-consistency rule for document URLs and a
tool-budget directive, and document the prompt style as a rule and guide.

Fix the find_links_matching browser tool, which double-encoded its
pattern and made JSON.parse fail on every keyword, starving any agent
that used it until it hit the turn cap. Salvage output when an agent
exhausts its turn budget while still exploring with a pending structured
output by forcing one final synthesis turn instead of failing outright.

Signed-off-by: Émile Ré <emile@probo.com>

2026-06-12 14:39:51 +02:00

2.2 KiB

Raw Permalink Blame History

Agent prompt style

Agent prompt templates (the //go:embed-ed *.txt.tmpl files that back an agent's instructions) follow a consistent XML-tag structure. This keeps prompts scannable, makes the contract between the prompt and the typed result struct explicit, and matches what the models in use respond to best.

Structure

Three top-level sections, always in this order:

<role>
One short paragraph: who the agent is and its single objective.
</role>

<task>
What the agent is given and the fields it must return. Describe the structured
output as a bullet list, one bullet per field, mirroring the typed result.
</task>

<instructions>
1. A numbered list of directives, ordered most-decisive first.
2. Each rule is a single, actionable directive.
3. ...
</instructions>

Rules

Use <role>, <task>, <instructions> in that order. Do not invent other top-level sections (no loose Method: / Rules: prose blocks).
<role> is one short paragraph stating who the agent is and its objective.
<task> enumerates the inputs and the output fields. List one bullet per field so the prompt stays in sync with the typed result struct.
<instructions> is an ordered (numbered) list. Put the most important rule first. Keep each item to a single directive.
Never hardcode enum values or source-of-truth lists. Use a {{.Placeholder}} and substitute at runtime. See file-naming.md and the template-files cursor rule.
No commentary outside the tags; the whole body is the system instruction.

Per-row input

The embedded template is the static system instruction. The dynamic per-row input (the specific entity being processed) is built separately in Go and sent as the user message, wrapped in its own descriptive tags:

fmt.Fprintf(&b, "<name> %s </name>\n", party.Name)
fmt.Fprintf(&b, "<website> %s </website>\n", website)

Examples

pkg/thirdparty/prompts/disambiguation.txt.tmpl
pkg/cookiebanner/prompts/tracker_identification.txt.tmpl
pkg/cookiebanner/prompts/tracker_enrichment.txt.tmpl
pkg/thirdparty/prompts/common_third_party_company_profile.txt.tmpl
pkg/thirdparty/prompts/common_third_party_compliance_docs.txt.tmpl

2.2 KiB Raw Permalink Blame History