Documentation
¶
Overview ¶
Package toolfmt provides a token-lean, information-preserving re-encoder for tool-call RESULTS before they enter an LLM's context. JSON is verbose (repeated keys, punctuation); EncodeTightC restates a uniform array of flat objects as a count-anchored table and drops needless quotes, cutting tokens for the same data while staying reversible.
Contract:
- Input that is NOT valid JSON (prose — file contents, `go test` output, plain text) is returned UNCHANGED. Only structured JSON is transformed.
- The transform is information-preserving and reversible: an LLM (or a decoder) reading the encoded form recovers every datum and its key/position.
JSON itself is the identity encoding — there is no EncodeJSON; the caller simply leaves Session.EncodeToolResult nil. tightc (EncodeTightC) is the one non-identity format, tuned for comprehension on small/quantized models.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func EncodeTightC ¶
EncodeTightC is TIGHT tuned for COMPREHENSION on small / quantized models, after the bonsai accuracy benchmark showed plain tight's token wins came at a real comprehension cost (aggregation/filtering/nested all degraded vs JSON and TOON). It keeps tight's cheap wins (bare tokens, unquoted-unless-needed, inline key:val objects, uniform tables) but spends a few tokens back where the data showed the model losing structure:
- ROW-COUNT ANCHOR on every table — `[N]col1,col2` header. bonsai answered "how many rows" questions with an EMPTY string on tight (no count to read) but correctly on TOON (which declares [N]). A count is not math the model computes; it is a structural anchor it reads. ~1 token per table.
- UNIFORM-ONLY tables. tight's SPARSE union table (`,,` empty cells for missing keys) forced the model to count commas to realign columns — the exact place filtering/aggregation misfired. Here a table is used ONLY when every row has every column; any irregularity drops to loose (explicit braces + keys on every value), which measured BETTER than tight on the nested dataset.
- NESTED OBJECTS render as loose JSON (explicit {}), not tab-indentation. Indentation is positional structure the model has to track; braces are explicit. loose beat tight by ~10 points on nested data.
Recoverability is unchanged (delimiters + header + [N] + loose braces). Non-JSON passes through. It is deliberately LESS token-minimal than tight — the bet is that on small models the comprehension gain outweighs the extra tokens.
Types ¶
This section is empty.
Source Files