Download the PHP package tivins/llm-php without Composer

Language: English

llm-php (tivins/llm-php)

Version: 1.21.1 (see CHANGELOG.md).

PHP client library for an OpenAI-compatible HTTP API—typically llama.cpp llama-server—covering POST /v1/chat/completions (non-stream and SSE stream), plus /health, /tokenize, and model discovery via GET /v1/models.

In this README: Install

Installation

Package

Requires: ext-curl.

llama.cpp server

• Install: llama.cpp install docs
• Server API: server README

Models

Use a GGUF checkpoint sized for your GPU/RAM; exact limits depend on quantization and context. The library does not download models—it only talks to a running server.

If you have a GPU with 8–16GB of VRAM, you can try these models:

• https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF (Q5_K_M or Q6_K)
• https://huggingface.co/bartowski/gemma-2-9b-it-GGUF (Q4_K_M)
• https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF (Q4_K_M)

Quick start

Prerequisite: start llama-server and keep it running so the PHP client can reach its OpenAI-compatible HTTP API. Example (adjust the GGUF path and port):

The snippet below assumes the default listen address http://127.0.0.1:8080.

From the repository root after composer install:

Lama::fromServerUrl() picks the first model id from /v1/models. Check $lama->isHealthy() returns true before relying on the server.

Sampling and generation knobs (temperature, top_p, max_tokens, penalties, seed, stop, n, tools, tool_choice) go through Tivins\Llama\ChatCompletionOptions: only properties you set are merged into the JSON body; omitted keys leave server defaults.

API surface: chat() vs chatCompletions() vs chatStream()

Fidelity path: use chatCompletions() and/or chatStream(), then map results into your own structures or into Tivins\Llama\Dto\NormalizedTurnOutcome (fromChatCompletionArray() / fromStreamResult()) for a single aggregate shape across modes.

Shortcut: chat() is documented in code as a thin wrapper around chatCompletions(); it is fine for plain text, not for tools or native reasoning traces.

Module map (src/Tivins/Llama/)

Client

• Tivins\Llama\Lama — HTTP client: fromServerUrl(), isHealthy() / getHealth(), tokenize(), chat(), chatCompletions(), chatStream().

Conversation model

• Tivins\Llama\Conversation — ordered Message list; toChatCompletionMessages() builds OpenAI-style messages (assistant tool_calls, tool role + tool_call_id, optional assistant reasoning_content).
• Tivins\Llama\Message — Role, content, optional toolCallId / name for Role::Tool, optional toolCalls (assistant), optional reasoningContent (native reasoning); normalizeReasoningContent() treats null and '' as absent for JSON.
• Tivins\Llama\Role — system, user, assistant, tool.

Request options

• Tivins\Llama\ChatCompletionOptions — OpenAI-shaped optional fields merged into the chat-completions body (see class docblock for semantics and local-server caveats).

Tools

• Tivins\Llama\ChatFunctionTool — build one tools[] entry (toToolArray(), toToolArrays()).
• Tivins\Llama\ToolCallingLoop — multi-round loop over chatCompletions(): executes tools, appends Role::Tool messages, copies reasoning_content when present; final assistant turn has no tool_calls when idle; throws if max rounds exhausted with pending tools.
• Tivins\Llama\StreamingToolCallingLoop — same orchestration over chatStream(); optional onAssistantStreamRound(StreamResult, RawStreamTrace, int) for logging (SSE capture when callback is used).
• Tivins\Llama\PredefinedTools — ready-made tools (search, fetch, filesystem helpers, apply_diff, git helpers, etc.) with executors suited to examples; see examples/ and class docblock (includes TLS-related environment variables). Paths are not confined to a project root by default.
• Tivins\Llama\WorkspacePaths / Tivins\Llama\WorkspaceToolExecutor — optional sandbox: resolve file_path, path, and working_directory under a workspace directory before calling PredefinedTools::runTool(). Pass WorkspaceToolExecutor::asCallable() to ToolCallingLoop as executeTool. See examples/workspace_tools_demo.php and examples/workspace_escape_demo.php (blocked access outside the root).

DTOs and audit

• Tivins\Llama\Dto\TurnRecord — one logical turn for JSONL (forCompletion / forStream, toLogArray()).
• Tivins\Llama\Dto\RawChatCompletionResponse — wraps non-stream completion JSON.
• Tivins\Llama\Dto\RawStreamTrace — events (list of StreamEvent) plus optional rawDataLines (verbatim SSE JSON strings).
• Tivins\Llama\Dto\StreamEvent / Tivins\Llama\Dto\StreamEventKind — structured stream replay types (fine-grained event lists may be empty depending on capture path).
• Tivins\Llama\Dto\NormalizedTurnOutcome — normalized assistant fields from completion JSON or StreamResult.
• Tivins\Llama\SsePayloadCapture — mutable bag of SSE JSON payload strings for RawStreamTrace::$rawDataLines.
• Tivins\Llama\TurnJsonlLogger — append one JSON line per TurnRecord.

Streaming aggregation

• Tivins\Llama\ChatStreamAccumulator — parses data: {...} SSE lines into StreamResult (shared by Lama::chatStream() and tests/fixtures).
• Tivins\Llama\StreamResult — content, finishReason, toolCalls, reasoningContent, optional usage, model, id.

Console rendering

• Tivins\Llama\RenderOptions — ANSI, stdout/stderr injectable, reasoning channel.
• Tivins\Llama\HumanTurnRenderer — render NormalizedTurnOutcome, TurnRecord, or raw completion payload (renderCompletionPayload()).
• Tivins\Llama\HumanTurnStreamDisplay — stream-friendly callbacks aligned with Lama::chatStream() / StreamingToolCallingLoop.

Higher-level helpers (optional)

• Tivins\Llama\ThinkingChat / ThinkingPrompts / ThinkingTurnResult — two HTTP rounds (reasoning prompt then answer); not the same as a single completion’s native reasoning_content (see class docblock).
• Tivins\Llama\BehaviorPrompts — ready-made system prompt strings.
• Tivins\Llama\Translator — translation helper built on Lama::chat() with optional FIFO cache.

Examples

Location: examples/. Scripts use:

Run from the repository root (so vendor/ resolves), with llama-server listening where the script expects (many examples use http://127.0.0.1:8080):

Additional demos include chat_tools.php, mediation.php, moderation.php, exemples.php, etc. Prefer reading each file’s header comment for prerequisites (e.g. patch on PATH for workspace_tools_demo.php).

Shared helpers: examples/_helpers.php (print_output(), JSONL helpers, render env parsing). Optional defaults: examples/.env (loaded without overriding variables already in the process environment).

Environment variables

Values are read via getenv() / putenv() in library or example code as documented below. Do not enable logging if completions could contain secrets.

Examples / console (examples/_helpers.php, examples/.env)

example_load_examples_env_file() reads examples/.env only for keys not already set in the environment.

HTTP/TLS for tool traffic (PredefinedTools, e.g. web_search / fetch_web_page)

With verification enabled, PredefinedTools sets CURLSSLOPT_NATIVE_CA when PHP 8.2+ provides it (OS trust store). On older PHP, configure curl.cainfo in php.ini if HTTPS tool calls fail.

Programmatic override: PredefinedTools::setHttpSslVerifyPeer() (documented on the class).

LangSearch web search (langsearch_web_search)

Programmatic override: PredefinedTools::setLangSearchApiKey(). Per PHP process, LangSearchRateLimiter enforces QPS 1, QPM 60, QPD 1000; when a per-minute or per-day quota is exhausted, further LangSearch tool calls in that session return a structured error with fallback_tool: web_search.

Tests

PHPUnit 11 (composer install then):

Unit tests live under tests/Unit/ (no live LLM server). Integration tests under tests/Integration/ (filesystem tools, git, optional network for web_search / fetch_web_page — group network).

Fixtures: tests/fixtures/.

Interactive / diagnostic:

• tests/stream_probe.php — live server: classifies incremental vs cumulative content deltas (complements fixture tests).

Conversation logging and modern message fields

Use this overview to wire audit logs, replay, and native reasoning without re-reading every subsection; details are in the linked sections.

Message and constructor compatibility

• Assistant messages support optional reasoningContent (wire key reasoning_content). New code should pass it with a named argument (reasoningContent: '…') so it does not collide with $toolCalls.
• Existing calls that only pass Role, content, and tool fields stay compatible. If you previously used five positional arguments for an assistant message, recall the signature is (role, content, toolCallId, name, toolCalls, reasoningContent) — add reasoning via the last named parameter rather than shifting arguments.

Tool-calling loops (behavior change)

• Since 1.14.0, ToolCallingLoop and StreamingToolCallingLoop append the final assistant turn (no tool_calls) when the model finishes, and throw if maxRounds is exhausted while tools are still pending. See CHANGELOG.md (1.14.0).

JSONL audit and replay

• Enable TIVINS_LLAMA_CONVERSATION_LOG in examples (Environment variables); each line is JSON from TurnRecord::toLogArray() (raw_completion or raw_stream + stream_result, optional request_messages for the prompt snapshot).
• Reconstruct records with TurnRecord::fromLogArray(); terminal replay: examples/replay_turn_jsonl.php. Full field list: JSONL audit logs.

Normalized view and console

• NormalizedTurnOutcome maps both non-stream completions and StreamResult into one shape (API surface).
• HumanTurnRenderer / HumanTurnStreamDisplay cover human-readable output (Console output).

Contributor-facing implementation history for these features lives in CHANGELOG.md (releases 1.14.0 through 1.20.x) and in the sections above.

JSONL audit logs (TurnJsonlLogger / TurnRecord)

• TurnJsonlLogger writes one JSON object per line from TurnRecord::toLogArray().
• Non-stream: TurnRecord::forCompletion() stores the full completion JSON under raw_completion.
• Stream: TurnRecord::forStream() includes StreamResult fields and RawStreamTrace; when SSE capture is used, raw_data_lines holds verbatim SSE JSON strings; structured events may be empty depending on the capture path.
• Request context: when examples pass request_messages (same shape as Conversation::toChatCompletionMessages()), logs include the prompt snapshot for that request; examples/replay_turn_jsonl.php prints it before options and assistant output.

Treat log files as sensitive if they can contain user data or downstream secrets.

Console output (HumanTurnRenderer / HumanTurnStreamDisplay)

• HumanTurnRenderer prints usage (if present), model/id, finish reason, reasoning block, content, and tool calls for NormalizedTurnOutcome or TurnRecord replay; renderCompletionPayload() adapts a raw chatCompletions() array.
• HumanTurnStreamDisplay separates streamed content (stdout by default), reasoning (stderr by default), and tool fragments / summaries (stderr), consistent with *`examples/stream_.php`**.
• On Windows, use a modern terminal (e.g. Windows Terminal) for ANSI; or set TIVINS_LLAMA_NO_ANSI.

Pitfalls and limits

• Lama::chat() discards everything except first-choice content—avoid it for tool calling or native reasoning_content.
• Streaming usage: many backends omit usage on SSE chunks; StreamResult::$usage stays null unless the server sends a usable usage object on a parsed chunk (the library keeps the last such object).
• Wire JSON shapes vary by server version; rely on this library’s aggregation helpers and tests for supported fields, or inspect raw payloads / logs.
• JSONL and captures can persist full prompts and completions—no secrets in shared logs.
• ChatCompletionOptions passes through keys the server may ignore or reject—check your backend’s supported subset.

License

MIT — see composer.json.