中文
← Back to Projects

Independent projectAI systems · Web

IndexWise

Research with context and sources.

Bring portfolio context, market data and an AI research assistant together while keeping the source and timing of claims visible.

227
commits
448
tests
35
tools for the adviser
4.98 / 5
grounding eval baseline

As of 2026-06-18, from the repository and releases

IndexWise — A starting point for research.
01A starting point for research.The current captured web entry screen. A research-session walkthrough is still to be added.

This capture shows the web sign-in screen. The research workspace described here is the project direction; the gallery does not yet show an authenticated session.

How it works

  1. 01

    Set the context

    The user brings a question and portfolio context into the Alfred interface.

  2. 02

    Ask Alfred

    Registered tools gather holdings, market data, analysis or external research.

  3. 03

    Read the evidence

    Tool output includes data, source and as-of fields; numeric claims are checked against available evidence.

  4. 04

    Explore further

    The answer arrives progressively with supporting context and unverified markers where needed.

The hard parts, and what I did

A research answer should show its working.

Keep a question, its portfolio context and the source of each claim close enough to examine together.

  1. An adviser that may not make up numbers

    The problem
    A language model will happily quote a fund’s return that it never looked up. In finance that is worse than no answer.
    What I did
    Every tool returns data with its source and time. A validator extracts the numbers in each answer and checks them against the tool results; anything it cannot match is marked “unverified”.
    Result
    Scored 4.98 / 5 on its own grounding evaluation (four dimensions, LLM-judged).
  2. Discipline where it matters

    The problem
    Letting a model rank funds makes the ranking impossible to audit.
    What I did
    Scoring and screening are a rule-based pipeline with no model in the loop; the adviser only explains what the pipeline found.
    Result
    A strategy backtest harness with costs included keeps the rules honest.

Independent projectIndependent research tool combining market data and an AI assistant; historical analysis is not a forecast.

Stack

React · TypeScript · ViteFastAPI · Pythonpandas · AKShare · httpxLLM tool registryGrounding validatorSSE response streamJWT authentication

Implementation details

Research software. Historical analysis does not promise future performance.

Research question
Start with a question grounded in a portfolio or market.
Research desk
Bring the dashboard and Alfred conversation together.
Alfred engine
Select tools and assemble an answer from their results.
Evidence checks
Keep data provenance and check numeric claims.
Market analysis
Transform data into historical comparisons and research context.
Data providers
Fetch configured market sources with their availability limits.