llmslim Files

LLMSlim Core 0.9.0: transcripts are selected turn by turn, agent histories fit a budget in gptme, Hermes Agent and Pydantic AI, and Sarvam, Exa and Neo4j integrations run behind a usage guard.

:::bash
pip install --upgrade llmslim

Changed: transcripts

• Plain-text transcripts are selected turn by turn (#48). A block of speaker-prefixed lines (Caroline: ..., user: ..., optionally after a stamp such as [8 May 2023]) is no longer joined into one paragraph and cut into sentence windows that straddle turns. Each turn is its own unit, a turn that asks a question stays with its reply, and a stamp shared by the whole block is shown once.
• Transcripts keep more of the evidence a question needs (#49). A reply that asks a question back keeps its own reply too, and in "how long / how many / how much" questions the word after "how" no longer matches text such as "long time no see".
• Measured once on round 3's test split, after both changes were chosen on dev, with Claude Haiku 5.5 reading:
• LoCoMo (1,307 questions): the default mode goes from 84.4% to 87.5% at 1,000 tokens, 90.9% to 91.7% at 2,500 and 94.0% to 94.3% at 10,000.
• LongMemEval-S (180 questions): 70.6% at 2,500 tokens and 77.8% at 10,000 (official judge), 1.7 points above [#48] alone.
• Dense retrieval over turns is still ahead on LongMemEval (76.1% and 80.6%). A further fix that narrowed that gap on dev did not hold up on test, so it is not in this release (#51, reverted in [#52]).
• Every answer, the cost per 1,000 questions and the weak spots: benchmarks/live/round3/RESULTS.md (github.com).

Added

• Agent-history adapters (#44). Each fits an agent's message history into a token budget locally, with no model call: the system prompt, tool calls and latest turns stay as they are, older tool output is compressed first, and tool calls stay paired with their results.
• gptme: llmslim.integrations.gptme.LLMSlimContextProvider; install llmslim[gptme].
• Hermes Agent: llmslim.integrations.hermes.LLMSlimContextEngine, selected with context.engine: llmslim.
• Pydantic AI: llmslim.integrations.pydantic_ai.llmslim_history_processor(max_tokens); install llmslim[pydantic-ai].
• Usage guard for paid integrations (UsageGuard). It enforces a per-minute rate, a lifetime request cap and an optional unit cap. Each call reserves before it leaves the process and settles with the vendor-reported amount.
• Sarvam, beyond chat (llmslim[sarvam]):
• an async provider;
• language identification, transliteration and translation;
• a PDF and image document source;
• profiles repriced to ₹15 / ₹5 / ₹60 per 1M input / cached / output tokens.
• Exa web search (llmslim[exa]): ExaSearchSource returns results as untrusted RAG items with URL, date and rank, capped in USD.
• Neo4j temporal agent memory (llmslim[neo4j]): Neo4jContextStore keeps facts with validity intervals and provides supersede(), recall(as_of=...), history() and conflicts().

Guides with runnable examples are in docs/integrations/.

Security

• Untrusted content can no longer close or forge the planner's <llmslim_context> provenance envelope. Trusted content renders byte for byte as before.
• Web-search results with unsafe URLs are dropped. The text sanitizer removes Unicode tag characters ("ASCII smuggling") and keeps ZWNJ and ZWJ.

Fixed

• Sentence centrality on large chunks. Above 4,096 sentences it no longer builds the full similarity matrix: a 24,000-sentence chunk takes about 150 MB instead of 2.3 GB. Thanks to @rlippmann (#17).
• Planner, Indic documents. A relevant one-sentence Indic document was cut mid-sentence (regression in 0.8.0). Planner results are again identical to 0.6.0 and 0.7.x on the v0.6 and v0.7 suites.
• Exa and Sarvam. Exa date and truncation fixes. Sarvam clients never follow redirects or retry inside the SDK, and they close cleanly.

Full changelog: CHANGELOG.md (github.com)