UNPKG

aios-core

Version:

Synkra AIOS: AI-Orchestrated System for Full Stack Development - Core Framework

58 lines (43 loc) 2.21 kB
--- paths: - .aios-core/data/tool-registry.yaml - .mcp.json --- # Tool Response Filtering — Dynamic Token Reduction When processing responses from MCP tools or large web fetches, apply the filter configuration defined in `.aios-core/data/tool-registry.yaml` for the tool that produced the response. This reduces context token consumption without losing task-relevant information. ## Filter Types ### content Extract the main informational content and discard noise (navigation, ads, boilerplate, repetitive headers/footers). Limit the extracted output to approximately `max_tokens` tokens, truncating at a natural paragraph or sentence boundary. If `extract` fields are specified, prioritize those fields from the response object. **Apply to:** WebFetch HTML responses, EXA search results, Context7 docs. ### schema From a JSON object or array of objects, select ONLY the fields listed in `fields`. Discard all other keys. If `max_tokens` is set, truncate the serialized result at that token limit. **Apply to:** Playwright page data, API responses with known schemas. ### field From an array of objects (tabular data), project ONLY the columns listed in `fields` and limit the result to `max_rows` rows. This is analogous to `SELECT field1, field2 FROM data LIMIT max_rows`. **Apply to:** Apify scraper results, database query results, CSV-like data. ## How to Apply 1. After receiving a tool response, identify the tool name. 2. Look up the tool in `tool-registry.yaml` → check for a `filter` key. 3. If a filter exists, apply the corresponding type rules above. 4. If NO filter exists for the tool, use the full response as-is. 5. Present the filtered result in your reasoning — do NOT repeat the raw unfiltered payload. ## Fallback If the filter would remove ALL content (empty result), fall back to the full unfiltered response. Never produce an empty result from filtering. ## Performance Note This is a zero-overhead optimization. The filter is applied during your reasoning step — no external scripts are invoked. The utility scripts at `.aios-core/utils/filters/` are available for batch post-processing of saved responses but are NOT required during normal tool use.