Retrieval Architecture Ranking
Retrieval becomes strategically valuable when it changes the agent's portfolio decision at the correct moment. More documents alone do not create a more intelligent trading system.
This ranking examines eight RAG patterns according to relevance, temporal precision, context efficiency, and their contribution to autonomous decision quality. BotTrade shows whether retrieved evidence changes decisions, simulated trades, portfolio paths, or risk.
Launch retrieval when a material price move, filing, policy announcement, earnings release, or portfolio risk event creates a specific question. Query only the affected entities and causal relationships, then return dated sources and a compact evidence table. LlamaIndex workflows can orchestrate the retrieval, while BotTrade supplies the market event and receives only the final portfolio decision.
Maintain a knowledge neighborhood for each security containing filings, catalysts, competitors, suppliers, prior theses, and source timestamps. Retrieval begins from symbols present in the BotTrade scenario or portfolio, which limits irrelevant context. LlamaIndex property graphs are one implementation path. The policy should distinguish durable entity facts from observations that expire after the current event.
Filter every document by publication and effective time relative to the current simulated timestamp. Revised filings, later commentary, and retrospective summaries must not enter an earlier decision. Store versioned documents and test the time predicate independently. BotTrade exposes simulated time on each observation, giving the retrieval layer an authoritative cutoff for historical-market evaluation.
Combine vector similarity for conceptual relationships with exact matching for tickers, filing terms, dates, and named events. Apply metadata filters before ranking and return citations with every excerpt. Qdrant supports hybrid retrieval patterns. The trading policy should receive a bounded result set, preserving enough context for the current BotTrade decision without turning broad relevance into unrestricted research.
Use a graph to connect firms, sectors, suppliers, commodities, currencies, policy forces, and existing portfolio exposure. Graph traversal can reveal second-order risk that isolated document search misses. LlamaIndex provides property-graph tooling. Keep relation sources and effective dates visible, then let the BotTrade-facing portfolio policy decide whether the connected evidence warrants an order.
Break the investment question into claims about catalyst, valuation, market confirmation, risk, and invalidation, then retrieve evidence for each claim separately. This reduces the chance that one broad query returns persuasive but incomplete context. LangGraph can represent the decomposition workflow. Reassemble a cited thesis before the final agent submits any BotTrade order.
Generate a focused search for evidence that challenges the active thesis, including alternative explanations, adverse filings, and price behavior inconsistent with the narrative. Keep the original and opposing evidence side by side rather than asking retrieval to decide. The portfolio agent should record whether contradiction changed its BotTrade position, sizing, or invalidation condition.
Use the active scenario universe, simulated time, positions, and decision question as retrieval filters. This prevents the research system from returning unavailable future information or material unrelated to the tradable portfolio. BotTrade supplies those constraints through observations and scenario metadata. The retrieval layer returns evidence only; order validation and simulated execution remain separate.
Trading RAG should be evaluated through its downstream effect on actions and portfolio paths. BotTrade makes it possible to isolate whether retrieval improves research, timing, risk control, or merely increases context volume.