NEWVectors or files. Pick a path.Start →
    Back to DiagramsAgentic Retrieval

    Search Ignores "Under 400 Dollars": The Agent Wrote the Filter Itself

    "I need to furnish a reading nook under 400 dollars": plain vector search returns 3 of 5 results over budget. Given the stage catalog as tools, the agent ran feature_search, then reached for attribute_filter and wrote price lt 400 itself. One result, $180, for $0.00061.

    A portrait diagram. The headline reads Search ignores, under 400 dollars, with the subline: an embedding has no idea what a budget is, so the agent picked a second stage and wrote the filter itself. A search bar holds the query I need to furnish a reading nook under 400 dollars, annotated: 3 of 5 results break the budget. A rail badged plain search, feature_search only, semantic match with no notion of price, shows five price cards: 922 dollars Steel Floor Lamp, 451 dollars Velvet Armchair and 663 dollars Leather Armchair outlined in rose as over budget, then 180 dollars Leather Armchair and 38 dollars Wood Floor Lamp; the chip reads 3 over budget. A bold line says the agent was given the same query and the whole stage catalog as tools. Two numbered trajectory cards follow: 1, feature_search with query furniture for reading nook, 18712 ms, 5 docs; 2, attribute_filter with price lt 400, 7 ms, 1 doc. A dark result band shows 180 dollars, Leather Armchair, the one result inside the budget, and the note that it derived price less than 400 from the sentence and nobody configured a filter. A footer line reads model gemini-2.5-flash-lite, 3 iterations, confidence 0.8, cost 0.00061 dollars, then mxp.co/d/search-that-reasons and real query, real trajectory, mixpeek.
    "I need to furnish a reading nook under 400 dollars": plain vector search returns 3 of 5 results over budget. Given the stage catalog as tools, the agent ran feature_search, then reached for attribute_filter and wrote price lt 400 itself. One result, $180, for $0.00061.

    An embedding has no representation of a number. Ask a 118 image index to furnish a reading nook under 400 dollars and plain vector search matches reading nook beautifully: three of the top five cost more than 400 dollars, at 922, 451 and 663.

    The second half of the diagram gives an agent the same sentence and the retriever's stage catalog as callable tools. Every step shown is read from the execution's own reasoning trace, tool call by tool call, with the real durations.

    Iteration one, the agent called feature_search with a rewritten query, furniture for reading nook, dropping the price clause it knew the embedding could not use. Five results in 18.7 seconds.

    Iteration two, it called a different stage, attribute_filter, and wrote the argument itself: price lt 400. Seven milliseconds. One result, a leather armchair at 180 dollars.

    Nobody configured that filter. The agent read the sentence, decided the budget belonged to a stage that understands numbers, and reached for it. Three iterations, confidence 0.8, at a cost of sixty one hundredths of a cent on gemini-2.5-flash-lite.

    The part worth sitting with is the second tool call. A language model parsing under 400 dollars is table stakes. A pipeline that exposes its stages as tools lets the model compose retrieval, filter and rerank steps on its own, per query, instead of running one fixed chain for every question.

    What is real here: the photos are Pexels images with every photographer credited, the embeddings and both executions ran against the live API, and the trajectory is the recorded one. Prices and categories are demo catalog data generated for this corpus.

    Run this on your own data

    Mixpeek turns video, images, audio, and documents in your object storage into searchable, timestamped results through one API.

    Search your own data

    Related diagrams

    Diagram contrasting a human query with an agent query, showing how agent calls carry accumulated state and prior tool output, and what a retrieval layer must return differently as a result.

    Agentic Retrieval: What Changes When the Caller Is Not a Person

    A portrait diagram in two rails. The headline reads Filters exclude, what ranking would have shown, with the subline: a pre_filter drops rows before ranking is ever shown to you, useful, and it hides what it removed. A search bar holds the query yellow armchair, annotated: three of the top five are out of stock. Rail A, badged no filter, shows five armchair photos in rank order with prices and scores: 872 dollars at 0.080, 515 dollars at 0.065, 180 dollars at 0.060, 451 dollars at 0.060 and 506 dollars at 0.057; the second, third and fourth are greyed and stamped dropped, and the count chip reads 5 of 5. Rail B, badged pre_filter with in_stock equals true, shows the two survivors at 872 and 506 dollars followed by three new armchairs at 123 dollars 0.048, 446 dollars 0.047 and 606 dollars 0.046; its chip reads 2 kept plus 3 new. A bold line states that three of the top five were out of stock and all three were removed from the response entirely. A dark code window shows the execute request with the filters line in_stock true highlighted, and skip_cache true. The footer prints mxp.co/d/filters-exclude and the words real query, real ranking, mixpeek.

    Filters Exclude What Ranking Would Have Shown

    A portrait diagram in two rails. The headline reads Scope decides, what can be found at all, with the subline: the same question, asked of the whole catalog and then of one category. The search bar holds the query something to keep me warm outside, annotated: the query names no category. Rail A, badged whole catalog with the note semantic reach across every category, shows five photos in rank order: a lit tent at night for 543 dollars at 0.062, a snowy forest scene for 564 dollars at 0.060, a steaming pot on a stove for 444 dollars at 0.059, boots on a textured rug for 203 dollars at 0.052, and a tent under a starry sky for 100 dollars at 0.048; the second, third and fourth are greyed and stamped dropped, and the chip reads 5 of 5. Rail B, badged scoped with category equals tent and the note only one category can answer, shows the two surviving tents followed by three more tents at 592 dollars 0.046, 73 dollars 0.039 and 554 dollars 0.029; its chip reads 2 kept plus 3 new. A bold line states that unscoped, the index answered with tents, backpacks, rugs and pans, and that scope is a decision about what is allowed to be right. A code window shows the execute request with the filters line category tent highlighted. The footer prints mxp.co/d/scope-changes-candidates.

    Scope Decides What Can Be Found at All