Add reranked_search tool

[Jack-Yu-815]

changes

1. added reranked_search as a new source package

   • the tool takes in sub-tools (other search tools), triggers them in parallel, and rerank based on returned content's relevance to query.
   • this reranker tool's input schema is dynamically created based on the input schemas of its sub-tools. Therefore, it can use sub-tools with different number of parameters, so long as they return some kind of search results. Usage is quite flexible.
   • sub-tool must utilize air_agent.common.SOURCE_DELIMITER to separate result strings.

2. small changes to other built-in tools in sources/

   • switched to using air_agent.common.SOURCE_DELIMITER to separate result strings.

3. updated dependency-related files and setup script.

[greptile-apps]

Greptile Summary

This PR introduces a reranked_search tool that fans out a query to multiple configured search tools in parallel and uses an NVIDIA cross-encoder (NVIDIARerank) to rerank the merged results, returning the top-k. It also standardises result separators across existing sources (tavily_web_search, knowledge_layer, paper_search) by introducing a shared SOURCE_DELIMITER constant.

• P1 — missing package dependency: sources/reranker/pyproject.toml imports from aiq_agent.common but does not declare aiq-agent as a dependency; standalone installs will fail with ModuleNotFoundError.
• P1 — header string is a rerank candidate: knowledge_layer._format_results prepends "Found N relevant document(s):\n" to the lines list that is later SOURCE_DELIMITER.join-ed, so the reranker treats this metadata header as a scoreable document, potentially displacing a real result.

Confidence Score: 3/5

Not safe to merge until the missing aiq-agent dependency and the header-as-rerank-candidate issues are resolved.

Two P1 defects remain: standalone installs of reranked-search will fail due to an undeclared aiq-agent dependency, and the knowledge_layer header string is incorrectly fed to the reranker as a scoreable candidate, silently degrading result quality.

sources/reranker/pyproject.toml (missing dependency) and sources/knowledge_layer/src/register.py (header in rerank pool)

Important Files Changed

Filename	Overview
sources/reranker/pyproject.toml	Missing aiq-agent dependency; package imports SOURCE_DELIMITER from aiq_agent.common but that package is not declared in dependencies.
sources/reranker/src/register.py	Core reranker implementation: fans out to sub-tools in parallel, reranks via NVIDIARerank; missing aiq-agent dep in pyproject.toml, empty strings passed to reranker, and dunder getattribute nit.
sources/knowledge_layer/src/register.py	Fixed SOURCE_DELIMITER chunking (per previous review), but the "Found N relevant document(s)" header is still included as the first rerank candidate via SOURCE_DELIMITER.join.
sources/tavily_web_search/src/register.py	Switched result separator to SOURCE_DELIMITER constant; change is clean and correct.
sources/google_scholar_paper_search/src/paper_search.py	Switched paper join separator to SOURCE_DELIMITER constant; straightforward and correct.
src/aiq_agent/common/prompt_utils.py	Added SOURCE_DELIMITER constant ("\n\n---\n\n") and exported it from common/init.py; clean change.
sources/reranker/README.md	New documentation for the reranked_search tool; describes configuration, environment variables, and compatibility requirements.
sources/reranker/example_cli_config.yml	Full working example config; prior llm field issue was flagged and acknowledged; config is otherwise representative.

Sequence Diagram

sequenceDiagram
    participant Agent
    participant RerankedSearch
    participant ToolA as search_tool_A
    participant ToolB as search_tool_B
    participant NVIDIARerank

    Agent->>RerankedSearch: _reranked_search(RerankedSearchInput)
    par Fan-out
        RerankedSearch->>ToolA: ainvoke(input.tool_a)
        ToolA-->>RerankedSearch: "result1\n\n---\n\nresult2"
    and
        RerankedSearch->>ToolB: ainvoke(input.tool_b)
        ToolB-->>RerankedSearch: "result3\n\n---\n\nresult4"
    end
    Note over RerankedSearch: split each by SOURCE_DELIMITER → all_results[]
    RerankedSearch->>NVIDIARerank: acompress_documents(query=overall_query, docs=all_results)
    NVIDIARerank-->>RerankedSearch: reranked_docs (scored)
    RerankedSearch-->>Agent: "Top k results ranked by relevance:\n..." 

Loading

_{Reviews (3): Last reviewed commit: "address greptile review comments, remove..." | Re-trigger Greptile}

[greptile-apps]

Missing aiq-agent dependency

sources/reranker/src/register.py imports from aiq_agent.common import SOURCE_DELIMITER, but aiq-agent is not listed in dependencies. This works in the uv workspace because every member can see the others, but running pip install reranked-search in any other environment will raise ModuleNotFoundError: No module named 'aiq_agent'. Add the internal package as a dependency:

dependencies = [
    "aiq-agent",
    "nvidia-nat==1.5.0",
    "pydantic>=2.0.0",
    "langchain-core>=1.1.0",
    "langchain-nvidia-ai-endpoints>=1.1.0"
]