> ## Documentation Index > Fetch the complete documentation index at: https://docs.anonalabs.com/llms.txt > Use this file to discover all available pages before exploring further. # Tuning Recall Quality > Retrieve or the drop-in proxy returning too much, too little, or the wrong memories? Here's what to change, and in what order. Retrieve and the drop-in proxy both ship with sensible defaults, but "sensible" isn't always right for your data. This guide is a symptom-to-fix map for the knobs that actually move recall quality. ## "It's returning duplicates of the same fact" This is by design, and there's a knob for it. Anona keeps two layers (raw `fact` memories and consolidated `note` memories synthesized from them), and by default (`prefer_observations: true`) a `note` replaces the facts it was built from, so you see one distilled result instead of the same content several times. If you're still seeing what looks like duplication, you're probably retrieving with `prefer_observations: false` (or an old space that predates consolidation catching up). Leave it at the default unless you specifically need the raw evidence behind a synthesis: ```python theme={null} results = client.retrieve( space_id="support", query="What plan are they on?", # prefer_observations defaults to True, so omit it ) ``` ## "It's returning too little" / "missing an obvious memory" Check these in order: 1. **Is `min_score` filtering it out?** `min_score` drops anything below the floor *without* shrinking `limit`, so a high floor can silently return fewer results than you asked for. Lower it or omit it. 2. **Is a scope key excluding it?** A scoped query (`user_id`/`agent_id`/`session_id`) returns *only* memories written under that exact scope, never unscoped ones, never another scope's. If the memory was written without a scope and you're querying with one, it won't come back. See [Scoping](/guides/scoping-multi-tenant). 3. **Is `limit` too low?** Default is 10. Raise it (`limit` / `top_k`, up to 100) if the memory exists but is being cut off by rank. 4. **Is it filtered by `memory_type` or `tags`?** Both narrow the result set; drop them to confirm the memory shows up unfiltered, then add them back deliberately. ## "It's returning too much" / "the wrong things are ranking high" 1. **Raise `min_score`.** This is the direct lever for precision vs. breadth. It trims low-relevance results without touching how many you asked for. 2. **Add `memory_type` or `tags` filters** if you know which kind of memory you actually want (e.g. `memory_type: ["fact"]` to skip synthesized notes and summaries). 3. **Check you're on `mode: "accurate"`, not `"fast"`.** `fast` skips the neural rerank pass and ranks on retrieval fusion alone: faster, but lower precision. It trades exactly the quality you're trying to improve. ## "The proxy is injecting too much / too little context into the model" This is the chat/responses/messages proxy, not raw `retrieve`, so the knobs are the memory tunables, not retrieve parameters: | Symptom | Fix | | ------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------- | | Prompt is bloated, costs are high | Lower `memory_limit` (default 5, max 20), or set `memory_token_budget` to cap the injected block by size instead of count. | | Model seems to be missing relevant context | Raise `memory_limit`. | | One call needs to skip memory entirely (e.g. a meta/system message) | `memory: false` for that call only. | | A read-only/eval call shouldn't pollute the space | `auto_record: false` for that call only. | Set these once as **space defaults** (`PUT /v1/spaces/{space_id}/chat-settings`) instead of a header on every call if the whole space should behave differently. Resolution order is `body → header → space default → platform default`, so a space default is a safe place to change behavior without touching call sites. See [Migrating to the drop-in proxy](/guides/migrating-to-proxy#step-3-tune-what-gets-injected-optional). ## "I want the model to see the memories exactly, not rely on the proxy's injection" Build the prompt yourself with `format: "block"` on `retrieve`, which returns a ready-to-paste string with the token budget already applied: ```json theme={null} { "space_id": "support", "query": "what plan are they on?", "format": "block", "context_max_tokens": 500 } ``` `context_max_tokens` drops **whole memories**, lowest-ranked first, until the block fits, and text is never truncated mid-sentence. `results` is still returned alongside the block, so you get both for one round trip. ## "Recall is too slow" Latency-sensitive path (an agent loop, autocomplete-style lookups)? Set `mode: "fast"` on `retrieve`, which skips the neural cross-encoder rerank and ranks on retrieval fusion alone. Accept the precision tradeoff deliberately, and don't set it globally as a default without checking result quality first. ## Quick reference | Parameter | Surface | Effect | | --------------------- | --------------------------------- | --------------------------------------------------------------------- | | `min_score` | `retrieve` | Precision floor; doesn't shrink `limit`. | | `limit` / `top_k` | `retrieve` | How many results, 1–100. | | `mode` | `retrieve` | `accurate` (default, reranked) vs. `fast` (fusion only). | | `prefer_observations` | `retrieve` | `true` (default) collapses facts into their note; `false` shows both. | | `memory_limit` | chat proxy | How many memories injected, 1–20. | | `memory_token_budget` | chat proxy | Token cap on the injected block. | | `context_max_tokens` | `retrieve` with `format: "block"` | Token cap on the returned block. | ## Next steps Full parameter and response reference. Every proxy tunable, body and header form.